shodh-memory 0.2.0

//! Memory System for LLM Context Management
//!
//! A medium-complexity memory system that provides:
//! - Hierarchical memory storage (working → session → long-term)
//! - Smart compression based on age and importance
//! - Multi-modal retrieval (similarity, temporal, causal)
//! - Automatic memory consolidation

pub mod compression;
pub mod context;
pub mod facts;
pub mod feedback;
pub mod files;
pub mod graph_retrieval;
pub mod hybrid_search;
pub mod injection;
pub mod introspection;
pub mod learning_history;
pub mod lineage;
pub mod pattern_detection;
pub mod prospective;
pub mod query_parser;
pub mod replay;
pub mod retrieval;
pub mod segmentation;
pub mod sessions;
pub mod storage;
pub mod temporal_facts;
pub mod todo_formatter;
pub mod todos;
pub mod types;
pub mod visualization;

use anyhow::{Context, Result};
use parking_lot::RwLock;
use serde::{Deserialize, Serialize};
use sha2::{Digest, Sha256};
use std::collections::HashSet;
use std::path::{Path, PathBuf};
use std::sync::Arc;
use tracing::debug;
use uuid::Uuid;

use crate::metrics::{
    EMBEDDING_CACHE_CONTENT, EMBEDDING_CACHE_CONTENT_SIZE, EMBEDDING_CACHE_QUERY,
    EMBEDDING_CACHE_QUERY_SIZE,
};

use crate::constants::{
    DEFAULT_COMPRESSION_AGE_DAYS, DEFAULT_IMPORTANCE_THRESHOLD, DEFAULT_MAX_HEAP_PER_USER_MB,
    DEFAULT_SESSION_MEMORY_SIZE_MB, DEFAULT_WORKING_MEMORY_SIZE, EDGE_SEMANTIC_WEIGHT_FLOOR,
    HEBBIAN_BOOST_HELPFUL, HEBBIAN_DECAY_MISLEADING, POTENTIATION_ACCESS_THRESHOLD,
    POTENTIATION_MAINTENANCE_BOOST, TIER_PROMOTION_SESSION_AGE_SECS,
    TIER_PROMOTION_SESSION_IMPORTANCE, TIER_PROMOTION_WORKING_AGE_SECS,
    TIER_PROMOTION_WORKING_IMPORTANCE,
};

use crate::memory::storage::{MemoryStorage, SearchCriteria};
pub use crate::memory::types::*;
// pub use crate::memory::vector_storage::{VectorIndexedMemoryStorage, StorageStats};  // Disabled
use crate::embeddings::Embedder;
use crate::memory::compression::CompressionPipeline;
pub use crate::memory::compression::{
    ConsolidationResult, FactType, SemanticConsolidator, SemanticFact,
};
pub use crate::memory::facts::{FactQueryResponse, FactStats, SemanticFactStore};
pub use crate::memory::feedback::{
    apply_context_pattern_signals, calculate_entity_flow, calculate_entity_overlap,
    detect_negative_keywords, extract_entities_simple, process_implicit_feedback,
    process_implicit_feedback_with_semantics, signal_from_entity_flow, ContextFingerprint,
    FeedbackMomentum, FeedbackStore, FeedbackStoreStats, PendingFeedback, PreviousContext,
    SignalRecord, SignalTrigger, SurfacedMemoryInfo, Trend,
};
pub use crate::memory::files::{FileMemoryStats, FileMemoryStore, IndexingResult};
pub use crate::memory::graph_retrieval::{
    calculate_density_weights, spreading_activation_retrieve, ActivatedMemory,
};
pub use crate::memory::hybrid_search::{
    BM25Index, HybridSearchConfig, HybridSearchEngine, HybridSearchResult, RRFusion,
};
pub use crate::memory::introspection::{
    AssociationChange, ConsolidationEvent, ConsolidationEventBuffer, ConsolidationReport,
    ConsolidationStats, EdgeFormationReason, FactChange, InterferenceEvent, InterferenceType,
    MemoryChange, PruningReason, ReplayEvent, ReportPeriod, StrengtheningReason,
};
pub use crate::memory::learning_history::{
    LearningEventType, LearningHistoryStore, LearningStats, LearningVelocity, StoredLearningEvent,
};
pub use crate::memory::lineage::{
    CausalRelation, InferenceConfig, LineageBranch, LineageEdge, LineageGraph, LineageSource,
    LineageStats, LineageTrace, TraceDirection,
};
pub use crate::memory::prospective::ProspectiveStore;
pub use crate::memory::replay::{
    InterferenceCheckResult, InterferenceDetector, InterferenceRecord, ReplayCandidate,
    ReplayCycleResult, ReplayManager,
};
use crate::memory::retrieval::RetrievalEngine;
pub use crate::memory::retrieval::{
    AnticipatoryPrefetch, IndexHealth, MemoryGraphStats, PrefetchContext, PrefetchReason,
    PrefetchResult, ReinforcementStats, RetrievalFeedback, RetrievalOutcome, TrackedRetrieval,
};
pub use crate::memory::segmentation::{
    AtomicMemory, DeduplicationEngine, DeduplicationResult, InputSource, SegmentationEngine,
};
pub use crate::memory::sessions::{
    Session, SessionEvent, SessionId, SessionStats, SessionStatus, SessionStore, SessionStoreStats,
    SessionSummary, TemporalContext, TimeOfDay,
};
pub use crate::memory::temporal_facts::{EventType, ResolvedTime, TemporalFact, TemporalFactStore};
pub use crate::memory::todos::{ProjectStats, TodoStore, UserTodoStats};
pub use crate::memory::visualization::{GraphStats, MemoryLogger};

/// Configuration for the memory system
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct MemoryConfig {
    /// Base directory for memory storage
    pub storage_path: PathBuf,

    /// Maximum size of working memory (in entries)
    pub working_memory_size: usize,

    /// Maximum size of session memory (in MB)
    pub session_memory_size_mb: usize,

    /// Maximum heap memory per user (in MB) - prevents OOM from single user
    pub max_heap_per_user_mb: usize,

    /// Enable auto-compression of old memories
    pub auto_compress: bool,

    /// Compression threshold (days)
    pub compression_age_days: u32,

    /// Importance threshold for long-term storage
    pub importance_threshold: f32,
}

impl Default for MemoryConfig {
    fn default() -> Self {
        Self {
            storage_path: PathBuf::from("./memory_store"),
            working_memory_size: DEFAULT_WORKING_MEMORY_SIZE,
            session_memory_size_mb: DEFAULT_SESSION_MEMORY_SIZE_MB,
            max_heap_per_user_mb: DEFAULT_MAX_HEAP_PER_USER_MB,
            auto_compress: true,
            compression_age_days: DEFAULT_COMPRESSION_AGE_DAYS,
            importance_threshold: DEFAULT_IMPORTANCE_THRESHOLD,
        }
    }
}

/// Main memory system
pub struct MemorySystem {
    config: MemoryConfig,

    /// Three-tier memory hierarchy
    working_memory: Arc<RwLock<WorkingMemory>>,
    session_memory: Arc<RwLock<SessionMemory>>,
    long_term_memory: Arc<MemoryStorage>,

    /// Compression pipeline
    compressor: CompressionPipeline,

    /// Retrieval engine
    retriever: RetrievalEngine,

    /// Embedder for semantic search
    embedder: Arc<crate::embeddings::minilm::MiniLMEmbedder>,

    /// Query embedding cache - SHA256(query_text) → embedding
    /// Uses SHA256 for stable hashing across restarts (unlike DefaultHasher)
    /// MASSIVE PERF WIN: 80ms → <1ms for cached queries
    /// LRU eviction: max 10,000 entries (~15MB for 384-dim embeddings)
    query_cache: moka::sync::Cache<[u8; 32], Vec<f32>>,

    /// Content embedding cache - SHA256(content) → embedding
    /// Uses SHA256 for stable hashing across restarts (unlike DefaultHasher)
    /// MASSIVE PERF WIN: 80ms → <1ms for repeated content
    /// LRU eviction: max 10,000 entries (~15MB for 384-dim embeddings)
    content_cache: moka::sync::Cache<[u8; 32], Vec<f32>>,

    /// Memory statistics
    stats: Arc<RwLock<MemoryStats>>,

    /// Visualization logger
    logger: Arc<RwLock<MemoryLogger>>,

    /// Consolidation event buffer for introspection
    /// Tracks what the memory system is learning (strengthening, decay, edges, facts)
    consolidation_events: Arc<RwLock<ConsolidationEventBuffer>>,

    /// Memory replay manager (SHO-105)
    /// Implements sleep-like consolidation through replay of high-value memories
    replay_manager: Arc<RwLock<replay::ReplayManager>>,

    /// Interference detector (SHO-106)
    /// Detects and handles memory interference (retroactive/proactive)
    interference_detector: Arc<RwLock<replay::InterferenceDetector>>,

    /// Pattern detector for intelligent replay triggers (PIPE-2)
    /// Replaces fixed 1-hour intervals with pattern-based consolidation
    pattern_detector: Arc<RwLock<pattern_detection::PatternDetector>>,

    /// Semantic fact store (SHO-f0e7)
    /// Stores distilled knowledge extracted from episodic memories
    /// Separate from episodic storage: facts persist, episodes flow
    fact_store: Arc<facts::SemanticFactStore>,

    /// Decision lineage graph (SHO-118)
    /// Tracks causal relationships between memories for "why" reasoning
    /// Enables: audit trails, project branching, automatic post-mortems
    lineage_graph: Arc<lineage::LineageGraph>,

    /// Hybrid search engine (BM25 + Vector + RRF + Reranking)
    /// Combines keyword matching with semantic similarity for better retrieval
    hybrid_search: Arc<hybrid_search::HybridSearchEngine>,

    /// Optional graph memory for entity relationships and spreading activation
    /// When set, entities are extracted and added to the knowledge graph on remember()
    /// This enables spreading activation retrieval and Hebbian co-activation learning
    graph_memory: Option<Arc<parking_lot::RwLock<crate::graph_memory::GraphMemory>>>,

    /// Optional feedback store for momentum-based scoring (PIPE-9)
    /// When set, retrieval applies feedback momentum to boost proven-helpful memories
    /// and suppress frequently-ignored memories (up to 20% penalty for negative momentum)
    feedback_store: Option<Arc<parking_lot::RwLock<FeedbackStore>>>,

    /// Persistent learning history for significant events
    /// Enables recency-weighted retrieval and learning velocity tracking
    learning_history: Arc<learning_history::LearningHistoryStore>,

    /// Temporal fact store for multi-hop temporal reasoning
    /// Extracts and indexes facts like "Melanie is planning camping next month"
    /// Resolves relative dates ("next month" → June 2023) for accurate retrieval
    temporal_fact_store: Arc<temporal_facts::TemporalFactStore>,

    /// Flag: new memories stored since last fact extraction cycle.
    /// When false, fact extraction is skipped entirely (no RocksDB scan, no clones).
    /// Set to true in remember(), cleared by maintenance after extraction runs.
    fact_extraction_needed: std::sync::atomic::AtomicBool,

    /// Watermark: only memories with created_at > this timestamp (unix millis) are
    /// processed for fact extraction. Persisted to RocksDB so server restarts don't
    /// re-process the entire memory store. Initialized from the latest fact's
    /// created_at or 0 if no facts exist.
    fact_extraction_watermark: std::sync::atomic::AtomicI64,

    /// Prediction cache for feedback prediction error weighting (VTA/Dopamine system).
    ///
    /// Stores (memory_id → final_score) for recently surfaced memories during recall.
    /// When feedback arrives (Helpful/Misleading), the prediction error (|predicted - actual|)
    /// scales the learning signal: expected outcomes → 0.5x, surprises → 2.0x.
    ///
    /// TTL: 10 minutes via moka cache. Cap: 500 entries (~4KB).
    ///
    /// Reference: Schultz et al. (1997) "A neural substrate of prediction and reward"
    prediction_cache: moka::sync::Cache<MemoryId, f32>,
}

/// Resolve an entity name to a graph label and salience using pre-extracted NER data.
///
/// Returns (EntityLabel, salience) based on NER type mapping, defaulting to (Concept, 0.5).
fn resolve_entity_label(
    entity_name: &str,
    ner_lookup: &std::collections::HashMap<String, (String, f32)>,
) -> (crate::graph_memory::EntityLabel, f32) {
    if let Some((ner_type, confidence)) = ner_lookup.get(&entity_name.to_lowercase()) {
        let label = match ner_type.as_str() {
            "PER" => crate::graph_memory::EntityLabel::Person,
            "ORG" => crate::graph_memory::EntityLabel::Organization,
            "LOC" => crate::graph_memory::EntityLabel::Location,
            _ => crate::graph_memory::EntityLabel::Concept,
        };
        (label, *confidence)
    } else {
        (crate::graph_memory::EntityLabel::Concept, 0.5)
    }
}

/// Build a lookup table from NER entity records for label resolution.
fn build_ner_lookup(
    ner_entities: &[NerEntityRecord],
) -> std::collections::HashMap<String, (String, f32)> {
    ner_entities
        .iter()
        .map(|r| (r.text.to_lowercase(), (r.entity_type.clone(), r.confidence)))
        .collect()
}

/// Tokenize text into word-level tokens for boundary-safe matching.
///
/// Splits on non-alphanumeric characters (preserving apostrophes for contractions),
/// filters tokens shorter than 3 characters. Returns borrowed slices to avoid allocation.
/// Caller should lowercase the input first for case-insensitive matching.
fn tokenize_words(text: &str) -> HashSet<&str> {
    text.split(|c: char| !c.is_alphanumeric() && c != '\'')
        .filter(|w| w.len() >= 3)
        .collect()
}

impl MemorySystem {
    /// Create a new memory system.
    ///
    /// If `shared_cache` is provided, all per-user RocksDB instances share the
    /// same LRU block cache (multi-tenant server mode). Pass `None` for
    /// standalone / test use — each DB gets a small local cache.
    pub fn new(config: MemoryConfig, shared_cache: Option<&rocksdb::Cache>) -> Result<Self> {
        let storage_path = config.storage_path.clone();
        let storage = Arc::new(
            MemoryStorage::new(&storage_path, shared_cache)
                .with_context(|| format!("Failed to open storage at {:?}", storage_path))?,
        );

        // CRITICAL: Initialize embedder ONCE and share between MemorySystem and RetrievalEngine
        // This prevents loading the ONNX model multiple times (50-200ms overhead per load)
        let embedding_config = crate::embeddings::minilm::EmbeddingConfig::default();
        let embedder = Arc::new(
            crate::embeddings::minilm::MiniLMEmbedder::new(embedding_config)
                .context("Failed to initialize MiniLM embedder (ONNX model)")?,
        );

        // Create consolidation event buffer first so we can share it with retriever
        let consolidation_events = Arc::new(RwLock::new(ConsolidationEventBuffer::new()));

        // Pass shared embedder and event buffer to retrieval engine (no duplicate model load)
        // Event buffer allows retriever to record Hebbian edge events for introspection
        let retriever = RetrievalEngine::with_event_buffer(
            storage.clone(),
            embedder.clone(),
            Some(consolidation_events.clone()),
        )
        .context("Failed to initialize retrieval engine")?;

        // STARTUP RECOVERY: Check for orphaned memories and auto-repair
        // This fixes memories that were stored but not indexed (crash, embedding failure, etc.)
        let storage_count = storage.get_stats().map(|s| s.total_count).unwrap_or(0);
        let indexed_count = retriever.len();
        let orphaned_count = storage_count.saturating_sub(indexed_count);

        if orphaned_count > 0 {
            tracing::warn!(
                storage_count = storage_count,
                indexed_count = indexed_count,
                orphaned_count = orphaned_count,
                "Detected orphaned memories at startup - initiating auto-repair"
            );

            // Get all memories from storage
            if let Ok(all_memories) = storage.get_all() {
                let indexed_ids = retriever.get_indexed_memory_ids();
                let mut repaired = 0;
                let mut failed = 0;

                for memory in all_memories {
                    if indexed_ids.contains(&memory.id) {
                        continue; // Already indexed
                    }

                    // Skip absurdly large memories (>1MB) - likely binary data or log dumps
                    // MiniLM only uses first ~512 tokens anyway, so this protects ONNX from hanging
                    const MAX_REPAIR_CONTENT_LEN: usize = 1_000_000;
                    if memory.experience.content.len() > MAX_REPAIR_CONTENT_LEN {
                        tracing::warn!(
                            memory_id = %memory.id.0,
                            content_len = memory.experience.content.len(),
                            "Skipping oversized memory during auto-repair (>1MB)"
                        );
                        failed += 1;
                        continue;
                    }

                    // Orphaned memory - try to index it
                    tracing::info!(memory_id = %memory.id.0, content_len = memory.experience.content.len(), "Attempting to repair orphaned memory...");
                    match retriever.index_memory(&memory) {
                        Ok(_) => {
                            repaired += 1;
                            if repaired <= 10 || repaired % 100 == 0 {
                                tracing::info!(
                                    memory_id = %memory.id.0,
                                    progress = format!("{}/{}", repaired, orphaned_count),
                                    "Repaired orphaned memory"
                                );
                            }
                        }
                        Err(e) => {
                            failed += 1;
                            tracing::error!(
                                memory_id = %memory.id.0,
                                error = %e,
                                "Failed to repair orphaned memory"
                            );
                        }
                    }
                }

                // Persist the repaired index
                if repaired > 0 {
                    if let Err(e) = retriever.save() {
                        tracing::error!("Failed to persist repaired index: {}", e);
                    } else {
                        tracing::info!(
                            repaired = repaired,
                            failed = failed,
                            final_indexed = retriever.len(),
                            "Startup repair complete - index persisted"
                        );
                    }
                }
            }
        } else if storage_count > 0 {
            tracing::info!(
                storage_count = storage_count,
                indexed_count = indexed_count,
                "All memories indexed - no repair needed"
            );
        }

        // Disable visualization logging for production performance
        let logger = Arc::new(RwLock::new(MemoryLogger::new(false)));

        // Load stats from storage to recover state after restart
        let initial_stats = {
            let storage_stats = storage.get_stats().unwrap_or_default();
            let vector_count = retriever.len();
            MemoryStats {
                total_memories: storage_stats.total_count,
                working_memory_count: 0, // Working memory is in-memory only, starts empty
                session_memory_count: 0, // Session memory is in-memory only, starts empty
                long_term_memory_count: storage_stats.total_count,
                vector_index_count: vector_count,
                compressed_count: storage_stats.compressed_count,
                promotions_to_session: 0,  // Runtime counter, not persisted
                promotions_to_longterm: 0, // Runtime counter, not persisted
                total_retrievals: storage_stats.total_retrievals,
                average_importance: storage_stats.average_importance,
                graph_nodes: 0, // Loaded separately from GraphMemory
                graph_edges: 0, // Loaded separately from GraphMemory
            }
        };

        // SHO-f0e7: Create semantic fact store using the same DB as long-term memory
        // Facts use "facts:" prefix to avoid key collisions with episodic memories
        let fact_store = Arc::new(facts::SemanticFactStore::new(storage.db()));

        // SHO-118: Create lineage graph for causal memory tracking
        // Lineage uses "lineage:" prefix for edges and branches
        let lineage_graph = Arc::new(lineage::LineageGraph::new(storage.db()));

        // Initialize hybrid search engine (BM25 + Vector + RRF + Reranking)
        let bm25_path = storage_path.join("bm25_index");
        let hybrid_search_config = hybrid_search::HybridSearchConfig::default();
        let hybrid_search_engine = hybrid_search::HybridSearchEngine::new(
            &bm25_path,
            embedder.clone(),
            hybrid_search_config,
        )
        .context("Failed to initialize hybrid search engine")?;

        // Backfill BM25 index if empty but memories exist
        if hybrid_search_engine.needs_backfill() {
            let existing_memories = storage.get_all()?;
            let memory_count = existing_memories.len();

            if memory_count > 0 {
                tracing::info!(
                    "BM25 index empty, backfilling {} existing memories...",
                    memory_count
                );

                let memories_iter = existing_memories.into_iter().map(|mem| {
                    (
                        mem.id,
                        mem.experience.content,
                        mem.experience.tags,
                        mem.experience.entities,
                    )
                });

                match hybrid_search_engine.backfill(memories_iter) {
                    Ok(indexed) => {
                        tracing::info!("BM25 backfill complete: {} memories indexed", indexed);
                    }
                    Err(e) => {
                        tracing::warn!("BM25 backfill failed (non-fatal): {}", e);
                    }
                }
            }
        }

        // Initialize learning history store for persistent significant events
        // Uses the same DB as long-term memory with "learning:" prefix
        let learning_history = Arc::new(learning_history::LearningHistoryStore::new(storage.db()));

        // Initialize temporal fact store for multi-hop temporal reasoning
        // Uses the same DB with "temporal_facts:", "temporal_by_entity:", "temporal_by_event:" prefixes
        let temporal_fact_store = Arc::new(temporal_facts::TemporalFactStore::new(storage.db()));

        // SHO-106: Load persisted interference history from RocksDB
        let interference_detector = {
            let mut detector = replay::InterferenceDetector::new();
            match storage.load_all_interference_records() {
                Ok((history, total_events)) => {
                    if !history.is_empty() {
                        detector.load_history(history, total_events);
                    }
                }
                Err(e) => {
                    tracing::warn!(
                        error = %e,
                        "Failed to load interference history, starting fresh"
                    );
                }
            }
            Arc::new(RwLock::new(detector))
        };

        Ok(Self {
            config: config.clone(),
            working_memory: Arc::new(RwLock::new(WorkingMemory::new(config.working_memory_size))),
            session_memory: Arc::new(RwLock::new(SessionMemory::new(
                config.session_memory_size_mb,
            ))),
            long_term_memory: storage,
            compressor: CompressionPipeline::new(),
            retriever,
            embedder,
            // LRU embedding caches: max 2,000 entries each (~3MB for 384-dim embeddings)
            query_cache: moka::sync::Cache::builder().max_capacity(2_000).build(),
            content_cache: moka::sync::Cache::builder().max_capacity(2_000).build(),
            stats: Arc::new(RwLock::new(initial_stats)),
            logger,
            consolidation_events, // Use the shared buffer created earlier
            // SHO-105: Memory replay manager
            replay_manager: Arc::new(RwLock::new(replay::ReplayManager::new())),
            // SHO-106: Interference detector (loaded from RocksDB)
            interference_detector,
            // PIPE-2: Pattern detector for intelligent replay triggers
            pattern_detector: Arc::new(RwLock::new(pattern_detection::PatternDetector::new())),
            // SHO-f0e7: Semantic fact store
            fact_store,
            // SHO-118: Decision lineage graph
            lineage_graph,
            // Hybrid search engine (always enabled)
            hybrid_search: Arc::new(hybrid_search_engine),
            // Graph memory is optional - wire up with set_graph_memory() for entity relationships
            graph_memory: None,
            // Feedback store is optional - wire up with set_feedback_store() for momentum scoring (PIPE-9)
            feedback_store: None,
            // Persistent learning history for retrieval boosting
            learning_history,
            // Temporal fact store for multi-hop temporal reasoning
            temporal_fact_store,
            // Dirty flag for fact extraction: run on first cycle, then only when new memories stored
            fact_extraction_needed: std::sync::atomic::AtomicBool::new(true),
            // Watermark for incremental fact extraction — initialized to 0 (sentinel).
            // On first maintenance call, loaded from RocksDB or derived from latest fact timestamp.
            fact_extraction_watermark: std::sync::atomic::AtomicI64::new(0),
            // Prediction cache for VTA/dopamine-inspired feedback error weighting
            // TTL 10 minutes, max 500 entries (~4KB)
            prediction_cache: moka::sync::Cache::builder()
                .max_capacity(500)
                .time_to_live(std::time::Duration::from_secs(600))
                .build(),
        })
    }

    /// Wire up GraphMemory for entity relationships and spreading activation
    ///
    /// When GraphMemory is set, the remember() method will:
    /// 1. Extract entities from memory content
    /// 2. Add them to the knowledge graph
    /// 3. Create edges between co-occurring entities
    ///
    /// This enables spreading activation retrieval and Hebbian learning
    pub fn set_graph_memory(
        &mut self,
        graph: Arc<parking_lot::RwLock<crate::graph_memory::GraphMemory>>,
    ) {
        self.graph_memory = Some(graph);
    }

    /// Get reference to the optional graph memory
    pub fn graph_memory(
        &self,
    ) -> Option<&Arc<parking_lot::RwLock<crate::graph_memory::GraphMemory>>> {
        self.graph_memory.as_ref()
    }

    /// Set the feedback store for momentum-based scoring (PIPE-9)
    ///
    /// When set, retrieval automatically applies feedback momentum:
    /// - Positive momentum (frequently helpful) → boost score
    /// - Negative momentum (frequently ignored) → suppress score (up to 20%)
    ///
    /// This provides consistent feedback integration across all retrieval paths.
    pub fn set_feedback_store(&mut self, feedback: Arc<parking_lot::RwLock<FeedbackStore>>) {
        self.feedback_store = Some(feedback);
    }

    /// Get reference to the optional feedback store
    pub fn feedback_store(&self) -> Option<&Arc<parking_lot::RwLock<FeedbackStore>>> {
        self.feedback_store.as_ref()
    }

    /// Store a new memory with an explicit ID.
    ///
    /// Used by MIF import to preserve original UUIDs. Stores the memory with
    /// embedding generation and vector indexing, but skips graph entity extraction
    /// (imported memories already have their entity relationships established).
    pub fn remember_with_id(
        &self,
        memory_id: MemoryId,
        mut experience: Experience,
        created_at: Option<chrono::DateTime<chrono::Utc>>,
    ) -> Result<MemoryId> {
        let importance = experience
            .importance_override
            .unwrap_or_else(|| self.calculate_importance(&experience));

        // Generate embedding if not provided
        if experience.embeddings.is_none() {
            let content_hash = Self::sha256_hash(&experience.content);
            if let Some(cached) = self.content_cache.get(&content_hash) {
                experience.embeddings = Some(cached.clone());
            } else if let Ok(embedding) = self.embedder.encode(&experience.content) {
                self.content_cache.insert(content_hash, embedding.clone());
                experience.embeddings = Some(embedding);
            }
        }

        let memory = Arc::new(Memory::new(
            memory_id.clone(),
            experience,
            importance,
            None,
            None,
            None,
            created_at,
        ));

        self.long_term_memory.store(&memory)?;
        self.logger.write().log_created(&memory, "import");

        self.working_memory
            .write()
            .add_shared(Arc::clone(&memory))?;

        if let Err(e) = self.retriever.index_memory(&memory) {
            tracing::warn!("Failed to index imported memory {}: {}", memory.id.0, e);
        }

        Ok(memory_id)
    }

    /// Store a new memory (takes ownership to avoid clones)
    /// Thread-safe: uses interior mutability for all internal state
    /// If `created_at` is None, uses current time (Utc::now())
    pub fn remember(
        &self,
        mut experience: Experience,
        created_at: Option<chrono::DateTime<chrono::Utc>>,
    ) -> Result<MemoryId> {
        // IDEMPOTENCY (issue #109): Check content hash index before creating a new memory.
        // If identical content already exists, return the existing MemoryId instead of
        // creating a duplicate. Catches all duplication paths: timeout retries, auto_ingest,
        // and manual re-remembers. O(1) RocksDB index lookup.
        if let Some(existing_id) = self
            .long_term_memory
            .get_by_content_hash(&experience.content)
        {
            tracing::debug!(
                existing_id = %existing_id.0,
                "Content dedup: returning existing memory (identical content already stored)"
            );
            return Ok(existing_id);
        }

        let memory_id = MemoryId(Uuid::new_v4());

        // Calculate importance
        let importance = experience
            .importance_override
            .unwrap_or_else(|| self.calculate_importance(&experience));

        // PERFORMANCE: Content embedding cache (80ms → <1μs for repeated content)
        // If experience doesn't have embeddings, check cache or generate
        if experience.embeddings.is_none() {
            // SHA256 hash for stable cache keys (survives restarts, unlike DefaultHasher)
            let content_hash = Self::sha256_hash(&experience.content);

            // Check cache first
            if let Some(cached_embedding) = self.content_cache.get(&content_hash) {
                experience.embeddings = Some(cached_embedding.clone());
                EMBEDDING_CACHE_CONTENT.with_label_values(&["hit"]).inc();
                tracing::debug!("Content embedding cache HIT");
            } else {
                // Cache miss - generate embedding
                EMBEDDING_CACHE_CONTENT.with_label_values(&["miss"]).inc();
                match self.embedder.encode(&experience.content) {
                    Ok(embedding) => {
                        // Store in cache for future use
                        self.content_cache.insert(content_hash, embedding.clone());
                        EMBEDDING_CACHE_CONTENT_SIZE.set(self.content_cache.entry_count() as i64);
                        experience.embeddings = Some(embedding);
                        tracing::debug!("Content embedding cache MISS - generated and cached");
                    }
                    Err(e) => {
                        tracing::warn!("Failed to generate embedding: {}", e);
                        // Continue without embedding - will be generated on-demand if needed
                    }
                }
            }
        }

        // TEMPORAL EXTRACTION: Extract dates from content for temporal filtering
        // Based on TEMPR approach (Hindsight paper achieving 89.6% on LoCoMo)
        if experience.temporal_refs.is_empty() {
            let temporal = crate::memory::query_parser::extract_temporal_refs(&experience.content);
            for temp_ref in temporal.refs {
                experience.temporal_refs.push(temp_ref.date.to_string());
            }
        }

        // Create memory entry (zero-copy with Arc)
        // CRITICAL: Move experience instead of clone to avoid 2-10KB allocation
        let memory = Arc::new(Memory::new(
            memory_id.clone(),
            experience, // Move ownership (zero-cost)
            importance,
            None,       // agent_id
            None,       // run_id
            None,       // actor_id
            created_at, // Use provided timestamp or Utc::now() if None
        ));

        // CRITICAL: Persist to RocksDB storage FIRST (before indexing/in-memory tiers)
        // This ensures retrieval can always fetch the memory from persistent storage
        self.long_term_memory.store(&memory)?;

        // Log creation
        self.logger.write().log_created(&memory, "working");

        // Add to working memory (cheap Arc clone, not full Memory clone)
        self.working_memory
            .write()
            .add_shared(Arc::clone(&memory))?;

        // CRITICAL: Index memory immediately for semantic search (don't wait for long-term promotion)
        // This ensures new memories are searchable right away, not only after consolidation
        let indexed = if let Err(e) = self.retriever.index_memory(&memory) {
            tracing::warn!("Failed to index memory {} in vector DB: {}", memory.id.0, e);
            // Don't fail the record operation if indexing fails - memory is still stored
            false
        } else {
            true
        };

        // NOTE: Graph processing (entities + co-occurrence edges) is handled by
        // process_experience_into_graph() at the handler layer (remember.rs, recall.rs).
        // That path creates richer EpisodicNodes with temporal context and does proper
        // entity embedding for concept-level dedup. Doing it here too would cause
        // double entity inserts and inflated mention_counts.

        // Index in BM25 for hybrid search (keyword + semantic)
        if let Err(e) = self.hybrid_search.index_memory(
            &memory.id,
            &memory.experience.content,
            &memory.experience.tags,
            &memory.experience.entities,
        ) {
            tracing::warn!("Failed to index memory {} in BM25: {}", memory.id.0, e);
        }

        // PIPE-2: Register memory for pattern-triggered replay
        // Tracks entity co-occurrence, salience spikes, and temporal clusters
        {
            let arousal = memory
                .experience
                .context
                .as_ref()
                .map(|c| c.emotional.arousal)
                .unwrap_or(0.0);

            let pattern_memory = pattern_detection::PatternMemory {
                id: memory.id.0.to_string(),
                content_preview: memory.experience.content.chars().take(100).collect(),
                entities: memory.experience.entities.clone(),
                importance,
                arousal,
                created_at: memory.created_at,
                embedding_hash: memory.experience.embeddings.as_ref().map(|e| {
                    e.iter()
                        .fold(0u64, |acc, &x| acc.wrapping_add(x.to_bits() as u64))
                }),
                session_id: memory
                    .experience
                    .context
                    .as_ref()
                    .and_then(|c| c.episode.episode_id.clone()),
                memory_type: format!("{:?}", memory.experience.experience_type),
            };

            let mut detector = self.pattern_detector.write();
            detector.register_memory(pattern_memory.clone());

            // Check for immediate salience spike (high-importance memory)
            if let Some(trigger) = detector.check_salience_spike(&pattern_memory) {
                tracing::debug!(
                    "Salience spike detected for memory {}: {}",
                    memory.id.0,
                    trigger.description()
                );
                // Record event for introspection
                self.record_consolidation_event(
                    introspection::ConsolidationEvent::PatternDetected {
                        trigger_type: trigger.trigger_type_name().to_string(),
                        description: trigger.description(),
                        memory_ids: trigger.memory_ids(),
                        timestamp: chrono::Utc::now(),
                    },
                );
            }
        }

        // TEMPORAL FACT EXTRACTION: Extract and index temporal facts for multi-hop reasoning
        // Key insight: Multi-hop temporal queries like "When is X planning Y?" require:
        // 1. Finding the FIRST/PLANNING mention, not any mention
        // 2. Resolving relative dates ("next month", "last Saturday") to absolute dates
        // This enables accurate answers to temporal questions in LoCoMo benchmark
        if !memory.experience.entities.is_empty() {
            let facts = temporal_facts::extract_temporal_facts(
                &memory.experience.content,
                &memory.id,
                memory.created_at,
                &memory.experience.entities,
            );
            if !facts.is_empty() {
                // Note: We don't have user_id in remember(), will need to pass it
                // For now, extract facts but don't store - storage happens at handler level
                // or we can use a placeholder user_id
                tracing::debug!(
                    "Extracted {} temporal facts from memory {}",
                    facts.len(),
                    memory.id.0
                );
            }
        }

        // SHO-106: Check for interference with existing memories
        // Find similar memories and apply retroactive/proactive interference
        if let Some(embedding) = &memory.experience.embeddings {
            // Search for similar memories (excluding the new one)
            if let Ok(similar_ids) =
                self.retriever
                    .search_by_embedding(embedding, 5, Some(&memory.id))
            {
                if !similar_ids.is_empty() {
                    // Collect similar memory data for interference check
                    let similar_memories: Vec<_> = similar_ids
                        .iter()
                        .filter_map(|(id, similarity)| {
                            self.retriever.get_from_storage(id).ok().map(|m| {
                                (
                                    id.0.to_string(),
                                    *similarity,
                                    m.importance(),
                                    m.created_at,
                                    m.experience.content.chars().take(50).collect::<String>(),
                                )
                            })
                        })
                        .collect();

                    if !similar_memories.is_empty() {
                        let interference_result =
                            self.interference_detector.write().check_interference(
                                &memory.id.0.to_string(),
                                importance,
                                memory.created_at,
                                &similar_memories,
                            );

                        // Apply retroactive interference (weaken old memories)
                        for (old_id, _similarity, decay_amount) in
                            &interference_result.retroactive_targets
                        {
                            if let Ok(old_uuid) = uuid::Uuid::parse_str(old_id) {
                                if let Ok(old_memory) =
                                    self.long_term_memory.get(&MemoryId(old_uuid))
                                {
                                    old_memory.decay_importance(*decay_amount);
                                    if let Err(e) = self.long_term_memory.update(&old_memory) {
                                        tracing::debug!("Failed to persist retroactive decay: {e}");
                                    }
                                }
                            } else {
                                tracing::warn!(
                                    "Skipping retroactive decay: malformed UUID '{old_id}'"
                                );
                            }
                        }

                        // Apply proactive interference (reduce new memory importance)
                        if interference_result.proactive_decay > 0.0 {
                            memory.decay_importance(interference_result.proactive_decay);
                            if let Err(e) = self.long_term_memory.update(&memory) {
                                tracing::debug!("Failed to persist proactive decay: {e}");
                            }
                        }

                        // Record interference events
                        for event in &interference_result.events {
                            self.record_consolidation_event(event.clone());
                        }

                        // Handle duplicates: suppress the near-duplicate to near-zero importance
                        // so it decays naturally. We can't delete it here because callers
                        // expect the returned MemoryId to be retrievable.
                        if interference_result.is_duplicate {
                            tracing::info!(
                                memory_id = %memory.id.0,
                                "Near-duplicate detected (≥0.95 cosine), suppressing importance"
                            );
                            // Heavy decay: drop to ~1% importance so natural decay removes it
                            memory.decay_importance(0.99);
                            if let Err(e) = self.long_term_memory.update(&memory) {
                                tracing::debug!("Failed to suppress duplicate importance: {e}");
                            }
                        }

                        // Persist affected interference records to RocksDB
                        {
                            let detector = self.interference_detector.read();
                            let affected_ids = detector.get_affected_ids_from_check(
                                &memory.id.0.to_string(),
                                &interference_result,
                            );
                            for (id, records) in detector.get_records_for_ids(&affected_ids) {
                                if let Err(e) =
                                    self.long_term_memory.save_interference_records(id, records)
                                {
                                    tracing::debug!("Failed to persist interference records: {e}");
                                }
                            }
                            let (total_events, _) = detector.stats();
                            if let Err(e) = self
                                .long_term_memory
                                .save_interference_event_count(total_events)
                            {
                                tracing::debug!("Failed to persist interference event count: {e}");
                            }
                        }
                    }
                }
            }
        }

        // If important enough, prepare for session storage
        let added_to_session = if importance > self.config.importance_threshold {
            self.session_memory
                .write()
                .add_shared(Arc::clone(&memory))?;
            self.logger.write().log_created(&memory, "session");
            true
        } else {
            false
        };

        // Update stats - track all tier counts accurately
        {
            let mut stats = self.stats.write();
            stats.total_memories += 1;
            stats.long_term_memory_count += 1; // Always stored to long-term first
            stats.working_memory_count += 1;
            if added_to_session {
                stats.session_memory_count += 1;
            }
            if indexed {
                stats.vector_index_count += 1;
            }
        }

        // Trigger background consolidation if needed
        self.consolidate_if_needed()?;

        // Commit and reload BM25 index changes (makes documents searchable immediately)
        // Note: This is done per-memory for immediate searchability.
        // For high-throughput scenarios, consider batching commits.
        if let Err(e) = self.hybrid_search.commit_and_reload() {
            tracing::warn!("Failed to commit/reload BM25 index: {}", e);
        }

        // Signal that fact extraction should run on next maintenance cycle
        self.fact_extraction_needed
            .store(true, std::sync::atomic::Ordering::Relaxed);

        Ok(memory_id)
    }

    /// Remember with agent context for multi-agent systems
    ///
    /// Same as `remember` but tracks which agent created the memory,
    /// enabling agent-specific retrieval and hierarchical memory tracking.
    pub fn remember_with_agent(
        &self,
        mut experience: Experience,
        created_at: Option<chrono::DateTime<chrono::Utc>>,
        agent_id: Option<String>,
        run_id: Option<String>,
    ) -> Result<MemoryId> {
        // IDEMPOTENCY (issue #109): Content hash dedup (same as remember())
        if let Some(existing_id) = self
            .long_term_memory
            .get_by_content_hash(&experience.content)
        {
            tracing::debug!(
                existing_id = %existing_id.0,
                "Content dedup: returning existing memory (identical content already stored)"
            );
            return Ok(existing_id);
        }

        let memory_id = MemoryId(Uuid::new_v4());

        // Calculate importance
        let importance = experience
            .importance_override
            .unwrap_or_else(|| self.calculate_importance(&experience));

        // PERFORMANCE: Content embedding cache
        if experience.embeddings.is_none() {
            let content_hash = Self::sha256_hash(&experience.content);
            if let Some(cached_embedding) = self.content_cache.get(&content_hash) {
                experience.embeddings = Some(cached_embedding.clone());
                EMBEDDING_CACHE_CONTENT.with_label_values(&["hit"]).inc();
            } else {
                EMBEDDING_CACHE_CONTENT.with_label_values(&["miss"]).inc();
                if let Ok(embedding) = self.embedder.encode(&experience.content) {
                    self.content_cache.insert(content_hash, embedding.clone());
                    EMBEDDING_CACHE_CONTENT_SIZE.set(self.content_cache.entry_count() as i64);
                    experience.embeddings = Some(embedding);
                }
            }
        }

        // TEMPORAL EXTRACTION: Extract dates from content for temporal filtering
        if experience.temporal_refs.is_empty() {
            let temporal = crate::memory::query_parser::extract_temporal_refs(&experience.content);
            for temp_ref in temporal.refs {
                experience.temporal_refs.push(temp_ref.date.to_string());
            }
        }

        // Create memory with agent context
        let memory = Arc::new(Memory::new(
            memory_id.clone(),
            experience,
            importance,
            agent_id,
            run_id,
            None, // actor_id
            created_at,
        ));

        // Persist to RocksDB storage
        self.long_term_memory.store(&memory)?;
        self.logger.write().log_created(&memory, "working");

        // Add to working memory
        self.working_memory
            .write()
            .add_shared(Arc::clone(&memory))?;

        // Index for semantic search
        if let Err(e) = self.retriever.index_memory(&memory) {
            tracing::warn!("Failed to index memory {} in vector DB: {}", memory.id.0, e);
        }

        // Add entities to knowledge graph with co-occurrence edges
        // PERF: Build entity structs and extract co-occurrences OUTSIDE the lock
        // GraphMemory is internally thread-safe; read lock allows concurrent graph access
        if let Some(graph) = &self.graph_memory {
            let now = chrono::Utc::now();

            // Phase 1: Build entity structs with proper labels from NER
            let ner_lookup = build_ner_lookup(&memory.experience.ner_entities);

            // Batch-encode entity names for concept-level dedup
            let entity_names: Vec<&str> = memory
                .experience
                .entities
                .iter()
                .map(|s| s.as_str())
                .collect();
            let entity_embeddings: Vec<Option<Vec<f32>>> = if entity_names.is_empty() {
                Vec::new()
            } else {
                match self.embedder.encode_batch(&entity_names) {
                    Ok(embs) => embs.into_iter().map(Some).collect(),
                    Err(e) => {
                        tracing::debug!(
                            error = %e,
                            "Entity name embedding failed, concept merge disabled for this batch"
                        );
                        vec![None; entity_names.len()]
                    }
                }
            };

            let entities_to_add: Vec<crate::graph_memory::EntityNode> = memory
                .experience
                .entities
                .iter()
                .zip(entity_embeddings.into_iter())
                .map(|(entity_name, embedding)| {
                    let (label, salience) = resolve_entity_label(entity_name, &ner_lookup);
                    crate::graph_memory::EntityNode {
                        uuid: Uuid::new_v4(),
                        name: entity_name.clone(),
                        labels: vec![label],
                        created_at: now,
                        last_seen_at: now,
                        mention_count: 1,
                        summary: String::new(),
                        attributes: std::collections::HashMap::new(),
                        name_embedding: embedding,
                        salience,
                        is_proper_noun: entity_name
                            .chars()
                            .next()
                            .map(|c| c.is_uppercase())
                            .unwrap_or(false),
                        selectivity: None,
                    }
                })
                .collect();

            // Phase 2: Use pre-extracted co-occurrence pairs or extract fresh
            let cooccurrence_pairs = if !memory.experience.cooccurrence_pairs.is_empty() {
                memory.experience.cooccurrence_pairs.clone()
            } else {
                let entity_extractor = crate::graph_memory::EntityExtractor::new();
                entity_extractor.extract_cooccurrence_pairs(&memory.experience.content)
            };

            let edge_context = format!("Co-occurred in memory {}", memory.id.0);

            // Phase 3: Acquire read lock for graph insertions (GraphMemory is internally thread-safe)
            let graph_guard = graph.read();

            for entity in entities_to_add {
                if let Err(e) = graph_guard.add_entity(entity.clone()) {
                    tracing::debug!("Failed to add entity '{}' to graph: {}", entity.name, e);
                }
            }

            // Semantic edge weighting
            let l1_base_weight = crate::graph_memory::EdgeTier::L1Working.initial_weight();
            for (entity1, entity2) in cooccurrence_pairs {
                if let (Ok(Some(e1)), Ok(Some(e2))) = (
                    graph_guard.find_entity_by_name(&entity1),
                    graph_guard.find_entity_by_name(&entity2),
                ) {
                    let entity_confidence = Some((e1.salience + e2.salience) / 2.0);

                    let semantic_weight = match (&e1.name_embedding, &e2.name_embedding) {
                        (Some(emb1), Some(emb2)) => {
                            let sim = crate::similarity::cosine_similarity(emb1, emb2).max(0.0);
                            EDGE_SEMANTIC_WEIGHT_FLOOR + (1.0 - EDGE_SEMANTIC_WEIGHT_FLOOR) * sim
                        }
                        _ => 1.0,
                    };

                    let edge = crate::graph_memory::RelationshipEdge {
                        uuid: Uuid::new_v4(),
                        from_entity: e1.uuid,
                        to_entity: e2.uuid,
                        relation_type: crate::graph_memory::RelationType::CoOccurs,
                        strength: l1_base_weight * semantic_weight,
                        created_at: now,
                        valid_at: now,
                        invalidated_at: None,
                        source_episode_id: Some(memory.id.0),
                        context: edge_context.clone(),
                        last_activated: now,
                        activation_count: 1,
                        ltp_status: crate::graph_memory::LtpStatus::None,
                        tier: crate::graph_memory::EdgeTier::L1Working,
                        activation_timestamps: None,
                        entity_confidence,
                        forman_curvature: None,
                        endpoint_selectivity: None,
                    };

                    if let Err(e) = graph_guard.add_relationship(edge) {
                        tracing::trace!(
                            "Failed to add co-occurrence edge {}<->{}: {}",
                            entity1,
                            entity2,
                            e
                        );
                    }
                }
            }
        }

        // Index in BM25 for hybrid search
        if let Err(e) = self.hybrid_search.index_memory(
            &memory.id,
            &memory.experience.content,
            &memory.experience.tags,
            &memory.experience.entities,
        ) {
            tracing::warn!("Failed to index memory {} in BM25: {}", memory.id.0, e);
        }

        // If important enough, add to session memory
        if importance > self.config.importance_threshold {
            self.session_memory
                .write()
                .add_shared(Arc::clone(&memory))?;
        }

        // Update stats
        {
            let mut stats = self.stats.write();
            stats.total_memories += 1;
            stats.long_term_memory_count += 1;
            stats.working_memory_count += 1;
        }

        self.consolidate_if_needed()?;

        // Commit and reload BM25 index changes
        if let Err(e) = self.hybrid_search.commit_and_reload() {
            tracing::warn!("Failed to commit/reload BM25 index: {}", e);
        }

        // Signal that fact extraction should run on next maintenance cycle
        self.fact_extraction_needed
            .store(true, std::sync::atomic::Ordering::Relaxed);

        Ok(memory_id)
    }

    /// Search and retrieve relevant memories (zero-copy with Arc<Memory>)
    ///
    /// PRODUCTION IMPLEMENTATION:
    /// - Semantic search: Uses embeddings + vector similarity across ALL tiers
    /// - Non-semantic search: Uses importance * temporal decay
    /// - Zero shortcuts, no TODOs, enterprise-grade
    pub fn recall(&self, query: &Query) -> Result<Vec<SharedMemory>> {
        // Semantic search requires special handling
        if let Some(query_text) = &query.query_text {
            return self.semantic_retrieve(query_text, query);
        }

        // Non-semantic search: filter-based retrieval
        let mut memories = Vec::new();
        let mut seen_ids: HashSet<MemoryId> = HashSet::new();
        let mut sources = Vec::new();

        // Collect from all tiers with deduplication (priority: working > session > long_term)
        {
            let working = self.working_memory.read();
            let working_results = working.search(query, query.max_results)?;
            if !working_results.is_empty() {
                sources.push("working");
            }
            for memory in working_results {
                if seen_ids.insert(memory.id.clone()) {
                    memories.push(memory);
                }
            }
        }

        {
            let session = self.session_memory.read();
            let session_results = session.search(query, query.max_results)?;
            if !session_results.is_empty() {
                sources.push("session");
            }
            for memory in session_results {
                if seen_ids.insert(memory.id.clone()) {
                    memories.push(memory);
                }
            }
        }

        {
            let long_term_results = self.retriever.search(query, query.max_results)?;
            if !long_term_results.is_empty() {
                sources.push("longterm");
            }
            for memory in long_term_results {
                if seen_ids.insert(memory.id.clone()) {
                    memories.push(memory);
                }
            }
        }

        // Expand with hierarchy context (parent chain + children)
        // Related memories in hierarchy get a decayed score boost
        self.expand_with_hierarchy(&mut memories, &mut seen_ids);

        // Rank by importance * temporal relevance
        let now = chrono::Utc::now();
        memories.sort_by(|a, b| {
            let age_days_a = (now - a.created_at).num_days();
            let temporal_a = Self::calculate_temporal_relevance(age_days_a);
            let score_a = a.importance() * temporal_a;

            let age_days_b = (now - b.created_at).num_days();
            let temporal_b = Self::calculate_temporal_relevance(age_days_b);
            let score_b = b.importance() * temporal_b;

            score_b.total_cmp(&score_a)
        });

        memories.truncate(query.max_results);

        // Log retrieval
        self.logger
            .read()
            .log_retrieved("", memories.len(), &sources);

        // Update access counts with instrumentation for consolidation events
        for memory in &memories {
            self.update_access_count_instrumented(memory, StrengtheningReason::Recalled);
        }

        // Hebbian learning: co-activation strengthens associations between memories
        // When memories are retrieved together, they form/strengthen edges in the memory graph
        if memories.len() >= 2 {
            if let Some(graph) = &self.graph_memory {
                let memory_uuids: Vec<uuid::Uuid> = memories.iter().map(|m| m.id.0).collect();
                if let Err(e) = graph.read().record_memory_coactivation(&memory_uuids) {
                    tracing::trace!("Coactivation recording failed (non-critical): {e}");
                }
            }
        }

        // Populate prediction cache for VTA/dopamine-inspired feedback error weighting.
        // Uses importance as the prediction signal — "how useful we think this memory is."
        // When feedback arrives later, the prediction error scales learning rate.
        for memory in &memories {
            self.prediction_cache
                .insert(memory.id.clone(), memory.importance());
        }

        // Increment and persist retrieval counter
        if let Ok(count) = self.long_term_memory.increment_retrieval_count() {
            self.stats.write().total_retrievals = count;
        }

        Ok(memories)
    }

    /// Paginated memory recall with "has_more" indicator (SHO-69)
    ///
    /// Returns a PaginatedResults struct containing:
    /// - The page of results
    /// - Whether there are more results beyond this page
    /// - The total count (if computed)
    /// - Pagination metadata (offset, limit)
    ///
    /// Uses the limit+1 trick: requests one extra result to detect if there are more.
    pub fn paginated_recall(&self, query: &Query) -> Result<PaginatedResults<SharedMemory>> {
        // Request offset+limit+1 to detect if there are more results.
        // We must fetch enough to cover both the skipped offset portion AND the
        // requested limit, plus 1 extra for has_more detection.
        let extra_limit = query.offset + query.max_results + 1;
        let mut modified_query = query.clone();
        modified_query.max_results = extra_limit;
        modified_query.offset = 0; // We handle offset ourselves

        // Get all results up to extra_limit
        let all_results = self.recall(&modified_query)?;

        // Apply offset and limit, detect has_more
        let offset = query.offset;
        let limit = query.max_results;

        let results_after_offset: Vec<_> = all_results.into_iter().skip(offset).collect();
        let has_more = results_after_offset.len() > limit;

        let final_results: Vec<_> = results_after_offset.into_iter().take(limit).collect();

        Ok(PaginatedResults {
            results: final_results,
            has_more,
            total_count: None, // Computing total would require a separate count query
            offset,
            limit,
        })
    }

    /// Recall memories by tags (fast, no embedding required)
    ///
    /// Returns memories that have ANY of the specified tags.
    pub fn recall_by_tags(&self, tags: &[String], limit: usize) -> Result<Vec<Memory>> {
        let criteria = storage::SearchCriteria::ByTags(tags.to_vec());
        let mut memories = self.advanced_search(criteria)?;
        memories.truncate(limit);
        if let Ok(count) = self.long_term_memory.increment_retrieval_count() {
            self.stats.write().total_retrievals = count;
        }
        Ok(memories)
    }

    /// Recall memories within a date range
    ///
    /// Returns memories created between `start` and `end` (inclusive).
    pub fn recall_by_date(
        &self,
        start: chrono::DateTime<chrono::Utc>,
        end: chrono::DateTime<chrono::Utc>,
        limit: usize,
    ) -> Result<Vec<Memory>> {
        let criteria = storage::SearchCriteria::ByDate { start, end };
        let mut memories = self.advanced_search(criteria)?;
        memories.truncate(limit);
        if let Ok(count) = self.long_term_memory.increment_retrieval_count() {
            self.stats.write().total_retrievals = count;
        }
        Ok(memories)
    }

    /// CACHE-AWARE semantic retrieval: Check working → session → storage
    ///
    /// Implementation:
    /// 1. Generate query embedding and search vector index for memory IDs
    /// 2. For each ID, check working memory (instant Arc clone)
    /// 3. If not found, check session memory (instant Arc clone)
    /// 4. Only fetch from RocksDB storage as last resort
    /// 5. This eliminates deserialization overhead for cached memories
    fn semantic_retrieve(&self, query_text: &str, query: &Query) -> Result<Vec<SharedMemory>> {
        let recall_start = std::time::Instant::now();

        // ===========================================================================
        // TEMPORAL ANALYSIS (unified intent + date parsing)
        // ===========================================================================
        // Replaces independent calls to extract_temporal_refs() and detect_temporal_intent().
        // analyze_temporal() parses dates first, then derives intent from parsed structure,
        // falling back to keyword detection only when dateparser finds nothing.
        let temporal_ctx = query_parser::analyze_temporal(query_text);
        let has_temporal_query = temporal_ctx.intent != query_parser::TemporalIntent::None;

        if has_temporal_query {
            tracing::debug!(
                "Temporal query detected: intent={:?}, filtering={}, seeking={}, date_range={:?}, refs={:?}",
                temporal_ctx.intent,
                temporal_ctx.is_filtering_query,
                temporal_ctx.is_seeking_query,
                temporal_ctx.date_range,
                temporal_ctx.extraction
                    .refs
                    .iter()
                    .map(|r| r.date.to_string())
                    .collect::<Vec<_>>()
            );
        }

        // ===========================================================================
        // LAYER 0.4: TEMPORAL PRE-FILTER (Date-Range Candidate Prefetch)
        // ===========================================================================
        // When the query has parsed temporal references (e.g., "yesterday", "March 2026"),
        // pre-fetch memories from the matching date range via SearchCriteria::ByDate.
        // These IDs are collected as candidates for boosting at Layer 4.45.
        // NOT exclusive — memories outside the range still enter via Vamana/BM25.
        let temporal_prefilter_ids: HashSet<MemoryId> = if temporal_ctx.is_filtering_query {
            if let Some((start_date, end_date)) = temporal_ctx.date_range {
                // Convert NaiveDate to DateTime<Utc> (start of first day, end of last day)
                let start_dt = start_date
                    .and_hms_opt(0, 0, 0)
                    .map(|dt| {
                        chrono::DateTime::<chrono::Utc>::from_naive_utc_and_offset(dt, chrono::Utc)
                    })
                    .unwrap_or_else(chrono::Utc::now);
                let end_dt = end_date
                    .and_hms_opt(23, 59, 59)
                    .map(|dt| {
                        chrono::DateTime::<chrono::Utc>::from_naive_utc_and_offset(dt, chrono::Utc)
                    })
                    .unwrap_or_else(chrono::Utc::now);

                match self.advanced_search(storage::SearchCriteria::ByDate {
                    start: start_dt,
                    end: end_dt,
                }) {
                    Ok(memories) => {
                        let ids: HashSet<_> = memories.iter().map(|m| m.id.clone()).collect();
                        if !ids.is_empty() {
                            tracing::info!(
                                    "Layer 0.4: Temporal pre-filter found {} memories in date range {:?} to {:?}",
                                    ids.len(),
                                    start_date,
                                    end_date
                                );
                        }
                        ids
                    }
                    Err(e) => {
                        tracing::warn!("Layer 0.4: Temporal pre-filter search failed: {}", e);
                        HashSet::new()
                    }
                }
            } else {
                HashSet::new()
            }
        } else {
            HashSet::new()
        };

        // ===========================================================================
        // LAYER 0.5: ATTRIBUTE QUERY DETECTION (Fact-First Retrieval)
        // ===========================================================================
        // For attribute queries like "What is Caroline's relationship status?",
        // semantic search fails because "relationship status" doesn't match "single".
        // Instead, we detect the query pattern, expand with synonyms, and boost
        // memories containing the entity + attribute values.
        let query_type = query_parser::classify_query(query_text);
        // Single parse for all layers (was called 3x: temporal facts, graph expansion, linguistic boost)
        let query_analysis = query_parser::analyze_query(query_text);

        // Ontological intent: infer expected entity types and relation types from query structure.
        // Used by Layer 2 (filtered traversal) and Layer 4.9 (type-aware re-ranking).
        let onto_intent = query_parser::infer_ontological_intent(query_text, &query_analysis);

        // Compute graph density for ontological gating (consistent with Layer 2 in graph_retrieval.rs).
        // Dense/young graphs have too many noisy L1 edges for type filtering to help.
        // Short-circuit: skip the RocksDB density lookups entirely when confidence is too low.
        let (graph_density_for_rerank, use_ontology_rerank) =
            if onto_intent.confidence < crate::constants::ONTOLOGICAL_MIN_CONFIDENCE {
                (None, false)
            } else if let Some(graph) = self.graph_memory.as_ref() {
                let g = graph.read();
                let seed_uuids: Vec<uuid::Uuid> = query_analysis
                    .focal_entities
                    .iter()
                    .filter_map(|e| {
                        g.find_entity_by_name(&e.text)
                            .ok()
                            .flatten()
                            .map(|n| n.uuid)
                    })
                    .collect();
                let density = if seed_uuids.is_empty() {
                    None
                } else {
                    g.entities_average_density(&seed_uuids).ok().flatten()
                };
                let use_rerank =
                    !density.is_some_and(|d| d >= crate::constants::ONTOLOGICAL_DENSITY_THRESHOLD);
                (density, use_rerank)
            } else {
                (None, false)
            };

        // Ontology telemetry
        crate::metrics::ONTOLOGICAL_INTENT_CONFIDENCE.observe(onto_intent.confidence as f64);
        if onto_intent.confidence > 0.0
            && onto_intent.confidence < crate::constants::ONTOLOGICAL_MIN_CONFIDENCE
        {
            crate::metrics::ONTOLOGICAL_FALLBACK_TOTAL.inc();
        }
        if graph_density_for_rerank
            .is_some_and(|d| d >= crate::constants::ONTOLOGICAL_DENSITY_THRESHOLD)
        {
            crate::metrics::ONTOLOGICAL_DENSITY_SKIP_TOTAL.inc();
        }

        let attribute_boost_ids: HashSet<MemoryId> = match &query_type {
            query_parser::QueryType::Attribute(attr_query) => {
                tracing::debug!(
                    "Layer 0.5: Attribute query detected - entity='{}', attribute='{}', synonyms={:?}",
                    attr_query.entity,
                    attr_query.attribute,
                    attr_query.attribute_synonyms
                );

                // Build expanded query: entity + attribute + all synonyms
                // E.g., "Caroline single married divorced engaged dating relationship"
                let mut expanded_terms: Vec<String> = vec![attr_query.entity.clone()];
                expanded_terms.extend(attr_query.attribute_synonyms.clone());

                // Search BM25 with expanded query to find memories with these terms
                let expanded_query = expanded_terms.join(" ");
                let bm25_matches = self
                    .hybrid_search
                    .bm25_index()
                    .search(&expanded_query, query.max_results * 5)
                    .unwrap_or_default();

                // Filter to memories that contain BOTH entity AND at least one synonym
                let entity_lower = attr_query.entity.to_lowercase();
                let mut boosted_ids = HashSet::new();

                for (mem_id, _score) in bm25_matches {
                    // Get memory content to verify it contains entity + attribute value
                    let content = self
                        .working_memory
                        .read()
                        .get(&mem_id)
                        .map(|m| m.experience.content.to_lowercase())
                        .or_else(|| {
                            self.session_memory
                                .read()
                                .get(&mem_id)
                                .map(|m| m.experience.content.to_lowercase())
                        })
                        .or_else(|| {
                            self.long_term_memory
                                .get(&mem_id)
                                .ok()
                                .map(|m| m.experience.content.to_lowercase())
                        });

                    if let Some(content) = content {
                        let content_words = tokenize_words(&content);
                        // Must contain all entity words (word-boundary safe)
                        if !entity_lower
                            .split_whitespace()
                            .all(|w| content_words.contains(w))
                        {
                            continue;
                        }
                        // Must contain at least one attribute synonym (word-level)
                        let has_synonym = attr_query.attribute_synonyms.iter().any(|syn| {
                            let syn_lower = syn.to_lowercase();
                            syn_lower
                                .split_whitespace()
                                .all(|w| content_words.contains(w))
                        });
                        if has_synonym {
                            boosted_ids.insert(mem_id);
                        }
                    }
                }

                if !boosted_ids.is_empty() {
                    tracing::info!(
                        "Layer 0.5: Found {} memories with entity '{}' + attribute values",
                        boosted_ids.len(),
                        attr_query.entity
                    );
                }

                boosted_ids
            }
            _ => HashSet::new(),
        };

        // ===========================================================================
        // LAYER 0.6: TEMPORAL FACT LOOKUP (Multi-hop Temporal Reasoning)
        // ===========================================================================
        // For temporal queries like "When did Melanie paint a sunrise?" or
        // "When is Melanie planning on going camping?", we need to:
        // 1. Detect it's a temporal query (asking "when", "what time", etc.)
        // 2. Extract entity (Melanie) and event keywords (paint, sunrise, camping)
        // 3. Look up temporal facts matching these
        // 4. Boost the source memories of matching facts
        // Temporal fact lookup - boost source memories of matching facts in Layer 4.5
        let temporal_fact_boost_ids: HashSet<MemoryId> = if has_temporal_query {
            if let Some(user_id) = &query.user_id {
                // Get entity name (first focal entity)
                let entity = query_analysis
                    .focal_entities
                    .first()
                    .map(|e| e.text.clone())
                    .unwrap_or_default();

                // Get event keywords from nouns, verbs, and modifiers
                let event_keywords: Vec<&str> = query_analysis
                    .focal_entities
                    .iter()
                    .skip(1) // Skip the entity itself
                    .map(|e| e.text.as_str())
                    .chain(
                        query_analysis
                            .relational_context
                            .iter()
                            .map(|r| r.stem.as_str()),
                    )
                    .chain(
                        query_analysis
                            .discriminative_modifiers
                            .iter()
                            .map(|m| m.text.as_str()),
                    )
                    .collect();

                if !entity.is_empty() && !event_keywords.is_empty() {
                    // Determine event type from query keywords
                    // "planning", "going to" → Planned
                    // "did", "ran", "went" → Occurred
                    // year mentions (2022, 2021) → Historical
                    let query_lower = query_text.to_lowercase();
                    let event_type = if query_lower.contains("planning")
                        || query_lower.contains("going to")
                        || query_lower.contains("will")
                    {
                        Some(temporal_facts::EventType::Planned)
                    } else if query_lower.contains(" did ")
                        || query_lower.contains("when did")
                        || query_lower.contains(" ran ")
                        || query_lower.contains(" went ")
                    {
                        // "When did X" could be Occurred or Historical - search both
                        None
                    } else {
                        None // Any event type
                    };

                    // Look up matching temporal facts
                    match self.find_temporal_facts(user_id, &entity, &event_keywords, event_type) {
                        Ok(facts) if !facts.is_empty() => {
                            tracing::info!(
                                "Layer 0.6: Found {} temporal facts for entity='{}', events={:?}",
                                facts.len(),
                                entity,
                                event_keywords
                            );

                            // Collect source memory IDs from matching facts
                            let boosted: HashSet<MemoryId> =
                                facts.iter().map(|f| f.source_memory_id.clone()).collect();
                            boosted
                        }
                        Ok(_) => {
                            tracing::debug!(
                                "Layer 0.6: No temporal facts found for entity='{}', events={:?}",
                                entity,
                                event_keywords
                            );
                            HashSet::new()
                        }
                        Err(e) => {
                            tracing::debug!("Layer 0.6: Temporal fact lookup failed: {}", e);
                            HashSet::new()
                        }
                    }
                } else {
                    HashSet::new()
                }
            } else {
                HashSet::new()
            }
        } else {
            HashSet::new()
        };

        // ===========================================================================
        // LAYER 0.7: SEMANTIC FACT SOURCE LOOKUP
        // ===========================================================================
        // Pre-fetch facts by query entities to boost their source memories in Layer 4.8.
        // Facts represent consolidated knowledge — their source memories contain the
        // richest context for that knowledge and should rank higher.
        let fact_source_boosts: std::collections::HashMap<MemoryId, f32> = {
            let mut boosts: std::collections::HashMap<MemoryId, f32> =
                std::collections::HashMap::new();

            if let Some(user_id) = &query.user_id {
                let entity_names: Vec<String> = query_analysis
                    .focal_entities
                    .iter()
                    .map(|e| e.text.to_lowercase())
                    .collect();

                if !entity_names.is_empty() {
                    if let Ok(facts) = self.get_facts_for_graph_entities(user_id, &entity_names, 5)
                    {
                        for fact in &facts {
                            if fact.confidence < 0.5 || fact.support_count < 3 {
                                continue;
                            }
                            let per_fact_boost = fact.confidence * 0.08;
                            for src_id in &fact.source_memories {
                                let entry = boosts.entry(src_id.clone()).or_insert(0.0);
                                *entry = (*entry + per_fact_boost).min(0.3);
                            }
                        }
                        if !boosts.is_empty() {
                            tracing::debug!(
                                "Layer 0.7: Pre-fetched {} fact-source boosts from {} facts",
                                boosts.len(),
                                facts.len()
                            );
                        }
                    }
                }
            }
            boosts
        };

        let t_query_analysis = recall_start.elapsed();
        tracing::info!(
            query_analysis_ms = format!("{:.2}", t_query_analysis.as_secs_f64() * 1000.0),
            "recall [layer:0.5-0.7] query analysis + attribute + temporal fact + fact source lookup"
        );

        // TEMPORAL PREFIX INJECTION: When the query has high-confidence temporal
        // references, prepend a temporal context string to give MiniLM textual
        // overlap with memory content mentioning the same time period.
        // Only applied for filtering queries (not "when did X happen?" seeking queries).
        let embedding_query_text: std::borrow::Cow<'_, str> = if temporal_ctx.is_filtering_query
            && temporal_ctx
                .extraction
                .refs
                .iter()
                .any(|r| r.confidence >= crate::constants::TEMPORAL_PREFIX_MIN_CONFIDENCE)
        {
            // Build prefix from the original temporal text (e.g., "March 2026", "yesterday")
            let temporal_labels: Vec<&str> = temporal_ctx
                .extraction
                .refs
                .iter()
                .filter(|r| r.confidence >= crate::constants::TEMPORAL_PREFIX_MIN_CONFIDENCE)
                .map(|r| r.original_text.as_str())
                .collect();
            let prefix = format!("[{}] ", temporal_labels.join(", "));
            tracing::debug!("Temporal prefix injection: \"{}\"", prefix);
            std::borrow::Cow::Owned(format!("{}{}", prefix, query_text))
        } else {
            std::borrow::Cow::Borrowed(query_text)
        };

        // PERFORMANCE: Use pre-computed embedding if caller provided one,
        // otherwise fall back to SHA256-keyed cache (80ms → <1μs for repeated queries)
        let query_embedding =
            if let Some(pre) = query.query_embedding.as_ref().filter(|e| !e.is_empty()) {
                EMBEDDING_CACHE_QUERY
                    .with_label_values(&["precomputed"])
                    .inc();
                tracing::debug!("Query embedding PRECOMPUTED by caller — skipping encode");
                pre.clone()
            } else {
                let query_hash = Self::sha256_hash(&embedding_query_text);
                if let Some(cached_embedding) = self.query_cache.get(&query_hash) {
                    EMBEDDING_CACHE_QUERY.with_label_values(&["hit"]).inc();
                    tracing::debug!("Query embedding cache HIT for: {}", &*embedding_query_text);
                    cached_embedding.clone()
                } else {
                    EMBEDDING_CACHE_QUERY.with_label_values(&["miss"]).inc();
                    tracing::debug!(
                        "Query embedding cache MISS - generating for: {}",
                        &*embedding_query_text
                    );
                    let embedding = self
                        .embedder
                        .as_ref()
                        .encode(&embedding_query_text)
                        .context("Failed to generate query embedding")?;

                    self.query_cache.insert(query_hash, embedding.clone());
                    EMBEDDING_CACHE_QUERY_SIZE.set(self.query_cache.entry_count() as i64);
                    embedding
                }
            };

        let t_embedding = recall_start.elapsed();
        tracing::info!(
            embedding_ms = format!(
                "{:.2}",
                (t_embedding - t_query_analysis).as_secs_f64() * 1000.0
            ),
            cumulative_ms = format!("{:.2}", t_embedding.as_secs_f64() * 1000.0),
            "recall [layer:embedding] query embedding (cache miss logged above if any)"
        );

        // ===========================================================================
        // LAYER 1: TEMPORAL PRE-FILTER (Episode Coherence)
        // ===========================================================================
        let episode_candidates: Option<HashSet<MemoryId>> = if let Some(episode_id) =
            &query.episode_id
        {
            match self
                .long_term_memory
                .search(SearchCriteria::ByEpisode(episode_id.clone()))
            {
                Ok(ep) if !ep.is_empty() => {
                    tracing::debug!("Layer 1: {} candidates in episode {}", ep.len(), episode_id);
                    Some(ep.into_iter().map(|m| m.id).collect())
                }
                _ => {
                    tracing::debug!("Layer 1: global search");
                    None
                }
            }
        } else {
            None
        };

        // ===========================================================================
        // LAYER 2: GRAPH EXPANSION (Knowledge Graph Traversal)
        // ===========================================================================
        let use_graph = matches!(
            query.retrieval_mode,
            RetrievalMode::Hybrid | RetrievalMode::Associative | RetrievalMode::Causal
        );
        let (
            graph_results,
            graph_density,
            query_entity_count,
            ic_weights,
            phrase_boosts,
            keyword_disc,
        ): (
            Vec<(MemoryId, f32, f32)>,
            Option<f32>,
            usize,
            std::collections::HashMap<String, f32>,
            Vec<(String, f32)>,
            f32, // Keyword discriminativeness for dynamic BM25/vector weight adjustment
        ) = {
            if let Some(graph) = self.graph_memory.as_ref().filter(|_| use_graph) {
                let g = graph.read();
                // Extract IC weights for BM25 term boosting
                let weights = query_analysis.to_ic_weights();
                // Extract phrase boosts for exact phrase matching (e.g., "support group")
                let phrases = query_analysis.to_phrase_boosts();
                // Extract keyword discriminativeness for dynamic weight adjustment
                // High discriminativeness → trust BM25 more for rare keywords like "sunrise"
                let (disc, disc_keywords) = query_analysis.keyword_discriminativeness();
                if disc > 0.5 && !disc_keywords.is_empty() {
                    tracing::debug!(
                        "Layer 2: YAKE discriminative keywords: {:?} (disc={:.2})",
                        disc_keywords,
                        disc
                    );
                }
                // Count entities in query for adaptive boost (multi-hop detection)
                let entity_count = query_analysis.focal_entities.len()
                    + query_analysis.discriminative_modifiers.len();

                // First, collect all query entity UUIDs
                // Include nouns, adjectives, AND verbs for multi-hop reasoning
                let mut query_entities: Vec<uuid::Uuid> = Vec::new();
                for e in query_analysis
                    .focal_entities
                    .iter()
                    .map(|e| e.text.as_str())
                    .chain(
                        query_analysis
                            .discriminative_modifiers
                            .iter()
                            .map(|m| m.text.as_str()),
                    )
                    .chain(
                        query_analysis
                            .relational_context
                            .iter()
                            .map(|r| r.text.as_str()),
                    )
                    .chain(
                        query_analysis
                            .relational_context
                            .iter()
                            .map(|r| r.stem.as_str()),
                    )
                {
                    if let Ok(Some(ent)) = g.find_entity_by_name(e) {
                        query_entities.push(ent.uuid);
                    }
                }

                // Calculate PER-ENTITY density (not global graph density)
                // Sparse entities = trust graph, Dense entities = trust vector
                let d = if !query_entities.is_empty() {
                    g.entities_average_density(&query_entities).ok().flatten()
                } else {
                    // No query entities — skip density calculation.
                    // The default weights (0.6, 0.3, 0.1) handle this case correctly.
                    None
                };

                let mut ids = Vec::new();

                // Density-adaptive traversal: dense graphs get shallower depth
                // and stricter strength filters to avoid exploring noisy L1 edges.
                // Dense graph results are already downweighted in RRF fusion
                // (graph_w=0.1 at density>2.0), so deep traversals add I/O cost
                // for results that contribute <0.01% to the fused score.
                let density_val = d.unwrap_or(0.0);
                let (bidir_depth, bidir_min_str, weighted_depth, weighted_min_str) =
                    if density_val > 15.0 {
                        (3usize, 0.12f32, 3usize, 0.15f32)
                    } else if density_val > 8.0 {
                        (4, 0.08, 4, 0.12)
                    } else {
                        (6, 0.05, 5, 0.10)
                    };

                if density_val > 0.0 {
                    tracing::debug!(
                        "Layer 2: density={:.1}, bidir_depth={}, bidir_min_str={:.2}, weighted_depth={}, weighted_min_str={:.2}",
                        density_val, bidir_depth, bidir_min_str, weighted_depth, weighted_min_str
                    );
                }

                // Multi-hop: Use bidirectional search between entity pairs
                // Cap to top 3 pairs from first 4 entities to avoid O(n²) explosion.
                // Entities are ordered by query analysis salience, so top pairs
                // capture dominant relationships.
                if query_entities.len() >= 2 {
                    let max_pairs = 3usize;
                    let max_ents = query_entities.len().min(4);
                    let mut pair_count = 0usize;
                    'bidir: for i in 0..max_ents {
                        for j in (i + 1)..max_ents {
                            if pair_count >= max_pairs {
                                break 'bidir;
                            }
                            if let Ok(path) = g.traverse_bidirectional(
                                &query_entities[i],
                                &query_entities[j],
                                bidir_depth,
                                bidir_min_str,
                            ) {
                                // Curvature-weighted path boost: positive curvature
                                // (community interior) increases boost, negative
                                // (bridge/bottleneck) decreases it.
                                let path_boost = {
                                    let curvatures: Vec<f32> = path
                                        .relationships
                                        .iter()
                                        .filter_map(|e| e.forman_curvature)
                                        .collect();
                                    if curvatures.is_empty() {
                                        1.5
                                    } else {
                                        let mean = curvatures.iter().sum::<f32>()
                                            / curvatures.len() as f32;
                                        (1.5 + mean * crate::constants::CURVATURE_PATH_BOOST_SCALE)
                                            .clamp(0.8, 2.5)
                                    }
                                };
                                for tr in &path.entities {
                                    if let Ok(mut eps) = g.get_episodes_by_entity(&tr.entity.uuid) {
                                        eps.sort_by(|a, b| b.created_at.cmp(&a.created_at));
                                        eps.truncate(50);
                                        for ep in eps {
                                            let mid = MemoryId(ep.uuid);
                                            if episode_candidates
                                                .as_ref()
                                                .map_or(true, |c| c.contains(&mid))
                                            {
                                                ids.push((
                                                    mid,
                                                    tr.entity.salience
                                                        * tr.decay_factor
                                                        * path_boost,
                                                    tr.decay_factor,
                                                ));
                                            }
                                        }
                                    }
                                }
                            }
                            pair_count += 1;
                        }
                    }
                }

                // Single-hop or supplement multi-hop: Weighted traversal from each entity.
                // When ontological intent has sufficient confidence, pass relation types
                // as a filter to traverse_weighted for type-aware graph expansion.
                let use_onto_filter = onto_intent.confidence
                    >= crate::constants::ONTOLOGICAL_MIN_CONFIDENCE
                    && !onto_intent.relation_types.is_empty();
                let relation_filter: Option<Vec<crate::graph_memory::RelationType>> =
                    if use_onto_filter {
                        Some(onto_intent.relation_types.clone())
                    } else {
                        None
                    };

                for entity_uuid in &query_entities {
                    if let Ok(t) = g.traverse_weighted(
                        entity_uuid,
                        weighted_depth,
                        relation_filter.as_deref(),
                        weighted_min_str,
                    ) {
                        let weighted_boost = {
                            let curvatures: Vec<f32> = t
                                .relationships
                                .iter()
                                .filter_map(|e| e.forman_curvature)
                                .collect();
                            if curvatures.is_empty() {
                                1.0
                            } else {
                                let mean = curvatures.iter().sum::<f32>() / curvatures.len() as f32;
                                (1.0 + mean * crate::constants::CURVATURE_PATH_BOOST_SCALE)
                                    .clamp(0.6, 2.0)
                            }
                        };
                        for tr in &t.entities {
                            if let Ok(mut eps) = g.get_episodes_by_entity(&tr.entity.uuid) {
                                eps.sort_by(|a, b| b.created_at.cmp(&a.created_at));
                                eps.truncate(50);
                                for ep in eps {
                                    let mid = MemoryId(ep.uuid);
                                    if episode_candidates
                                        .as_ref()
                                        .map_or(true, |c| c.contains(&mid))
                                    {
                                        ids.push((
                                            mid,
                                            tr.entity.salience * tr.decay_factor * weighted_boost,
                                            tr.decay_factor,
                                        ));
                                    }
                                }
                            }
                        }
                    }
                }

                let mut seen: std::collections::HashMap<MemoryId, (f32, f32)> =
                    std::collections::HashMap::new();
                for (id, act, heb) in ids {
                    seen.entry(id)
                        .and_modify(|(a, h)| {
                            *a = a.max(act);
                            *h = h.max(heb);
                        })
                        .or_insert((act, heb));
                }
                let mut r: Vec<_> = seen.into_iter().map(|(id, (a, h))| (id, a, h)).collect();
                // CRITICAL: Sort by activation score so RRF rank is meaningful
                r.sort_by(|a, b| b.1.total_cmp(&a.1));
                let pre_cap = r.len();
                // Cap total graph candidates to prevent flooding RRF fusion
                r.truncate(200);
                if !r.is_empty() {
                    tracing::debug!("Layer 2: {} graph results (capped from {}), {} query entities, bidirectional={}, top_activation={:.3}",
                        r.len(), pre_cap, entity_count, query_entities.len() >= 2, r.first().map(|x| x.1).unwrap_or(0.0));
                }
                (r, d, entity_count, weights, phrases, disc)
            } else {
                if !use_graph && self.graph_memory.is_some() {
                    tracing::debug!(
                        "Layer 2: SKIPPED (retrieval_mode={:?})",
                        query.retrieval_mode
                    );
                }
                // No graph traversal - still analyze query for IC weights and phrase boosts
                let (disc, _) = query_analysis.keyword_discriminativeness();
                (
                    Vec::new(),
                    None,
                    0,
                    query_analysis.to_ic_weights(),
                    query_analysis.to_phrase_boosts(),
                    disc,
                )
            }
        };

        let t_graph = recall_start.elapsed();
        tracing::info!(
            graph_ms = format!("{:.2}", (t_graph - t_embedding).as_secs_f64() * 1000.0),
            cumulative_ms = format!("{:.2}", t_graph.as_secs_f64() * 1000.0),
            graph_results = graph_results.len(),
            "recall [layer:1-2] episode filter + graph expansion"
        );

        // Create a modified query with the embedding for vector search
        let vector_query = Query {
            user_id: query.user_id.clone(),
            query_text: None, // Don't re-generate embedding
            query_embedding: Some(query_embedding),
            time_range: query.time_range,
            experience_types: query.experience_types.clone(),
            importance_threshold: query.importance_threshold,
            max_results: query.max_results,
            retrieval_mode: query.retrieval_mode.clone(),
            // Robotics filters (carry over from original query)
            robot_id: query.robot_id.clone(),
            mission_id: query.mission_id.clone(),
            geo_filter: query.geo_filter.clone(),
            action_type: query.action_type.clone(),
            reward_range: query.reward_range,
            // Decision & Learning filters (carry over from original query)
            outcome_type: query.outcome_type.clone(),
            failures_only: query.failures_only,
            anomalies_only: query.anomalies_only,
            severity: query.severity.clone(),
            tags: query.tags.clone(),
            pattern_id: query.pattern_id.clone(),
            terrain_type: query.terrain_type.clone(),
            confidence_range: query.confidence_range,
            offset: query.offset,
            episode_id: query.episode_id.clone(),
            session_id: query.session_id.clone(),
            prospective_signals: query.prospective_signals.clone(),
            recency_weight: query.recency_weight,
        };

        // ===========================================================================
        // LAYER 3: VECTOR SEARCH (Vamana Index)
        // ===========================================================================
        let vr = self
            .retriever
            .search_ids(&vector_query, query.max_results * 3)?;
        let vector_results: Vec<(MemoryId, f32)> = if let Some(ref c) = episode_candidates {
            vr.into_iter().filter(|(id, _)| c.contains(id)).collect()
        } else {
            vr
        };
        let t_vector = recall_start.elapsed();
        tracing::info!(
            vector_ms = format!("{:.2}", (t_vector - t_graph).as_secs_f64() * 1000.0),
            cumulative_ms = format!("{:.2}", t_vector.as_secs_f64() * 1000.0),
            vector_results = vector_results.len(),
            "recall [layer:3] Vamana vector search"
        );

        // ===========================================================================
        // LAYER 4: BM25 + RRF FUSION
        // ===========================================================================
        let (memory_ids, hebbian_scores): (
            Vec<(MemoryId, f32)>,
            std::collections::HashMap<MemoryId, f32>,
        ) = {
            let get_content = |id: &MemoryId| -> Option<String> {
                self.working_memory
                    .read()
                    .get(id)
                    .map(|m| m.experience.content.clone())
                    .or_else(|| {
                        self.session_memory
                            .read()
                            .get(id)
                            .map(|m| m.experience.content.clone())
                    })
                    .or_else(|| {
                        self.long_term_memory
                            .get(id)
                            .ok()
                            .map(|m| m.experience.content.clone())
                    })
            };
            // Use IC-weighted BM25 search with phrase matching
            let term_weights = if ic_weights.is_empty() {
                None
            } else {
                Some(&ic_weights)
            };
            let phrases = if phrase_boosts.is_empty() {
                None
            } else {
                Some(phrase_boosts.as_slice())
            };
            // Use dynamic weight adjustment based on YAKE keyword discriminativeness
            // High discriminativeness → boost BM25 weight for rare keywords
            let disc_opt = if keyword_disc > 0.3 {
                Some(keyword_disc)
            } else {
                None
            };
            let hybrid_ids = self
                .hybrid_search
                .search_with_dynamic_weights(
                    query_text,
                    vector_results.clone(),
                    get_content,
                    term_weights,
                    phrases,
                    disc_opt,
                )
                .map(|r| {
                    r.into_iter()
                        .map(|x| (x.memory_id, x.score))
                        .collect::<Vec<_>>()
                })
                .unwrap_or(vector_results);

            // ===========================================================================
            // LAYER 4: RRF FUSION WITH DENSITY-BASED WEIGHTS (PIPE-11)
            // ===========================================================================
            // Biological model: Memory graphs start dense (noisy L1 edges) and become
            // sparse over time through pruning (Hebbian "use it or lose it").
            //
            // Sparse graphs = mature, curated connections = trust graph more
            // Dense graphs = fresh, noisy connections = trust semantic/BM25 more
            //
            // The density weights directly control the balance - no extra multipliers.
            // This follows ACT-R's additive activation model.
            // K value from constants.rs — see Cormack et al. (2009), Anderson & Lebiere (1998)
            use crate::constants::RRF_K_GRAPH_FUSION;
            let k = RRF_K_GRAPH_FUSION;
            let mut fused: std::collections::HashMap<MemoryId, f32> =
                std::collections::HashMap::new();
            let mut heb: std::collections::HashMap<MemoryId, f32> =
                std::collections::HashMap::new();

            // Density-based weights (already tuned in calculate_density_weights)
            // Sparse (≤0.5): graph_w=0.5, semantic_w=0.4, linguistic_w=0.1
            // Dense (≥2.0):  graph_w=0.1, semantic_w=0.7, linguistic_w=0.2
            let (semantic_w, graph_w, linguistic_w) = graph_density
                .map(calculate_density_weights)
                .unwrap_or((0.6, 0.3, 0.1));

            // Hybrid weight = semantic + linguistic (BM25 + vector combined)
            let hybrid_w = semantic_w + linguistic_w;

            tracing::debug!(
                "Layer 4 RRF: density={:?}, graph_w={:.2}, hybrid_w={:.2}, query_entities={}",
                graph_density,
                graph_w,
                hybrid_w,
                query_entity_count
            );

            // Graph results: pure RRF with density weight
            for (r, (id, activation, h)) in graph_results.iter().enumerate() {
                // Standard RRF: weight / (k + rank), rank is 1-indexed
                let rrf_score = graph_w / (k + (r + 1) as f32);
                *fused.entry(id.clone()).or_insert(0.0) += rrf_score;
                heb.insert(id.clone(), *h);

                // Multiplicative activation bonus (ACT-R style spreading activation)
                // Scaled by graph_w: trust activation more when graph is sparse/mature
                let activation_factor = 1.0
                    + graph_w
                        * crate::constants::ACTIVATION_BONUS_SCALE
                        * activation.clamp(0.0, 1.0);
                *fused.get_mut(id).unwrap() *= activation_factor;
            }

            // Hybrid (BM25+vector) results: pure RRF with density weight
            for (r, (id, _)) in hybrid_ids.iter().enumerate() {
                *fused.entry(id.clone()).or_insert(0.0) += hybrid_w / (k + (r + 1) as f32);
            }

            // ===========================================================================
            // LAYER 4.5: ATTRIBUTE QUERY BOOST
            // ===========================================================================
            // For attribute queries, heavily boost memories that contain BOTH the entity
            // AND an attribute synonym value. This ensures "Caroline is single" ranks
            // high for "What is Caroline's relationship status?".
            if !attribute_boost_ids.is_empty() {
                let mut boosted_count = 0;
                for id in &attribute_boost_ids {
                    if let Some(score) = fused.get_mut(id) {
                        // Multiplicative: preserve base ranking while amplifying attribute matches
                        *score *= 1.0 + crate::constants::ATTRIBUTE_QUERY_BOOST;
                        boosted_count += 1;
                    }
                    // Don't insert memories absent from fusion — attribute match alone
                    // without semantic/graph support is insufficient evidence
                }
                if boosted_count > 0 {
                    tracing::info!(
                        "Layer 4.5: Boosted {} memories for attribute query",
                        boosted_count
                    );
                }
            }

            // ===========================================================================
            // LAYER 4.45: TEMPORAL PRE-FILTER BOOST
            // ===========================================================================
            // Memories that fell within the query's parsed date range (Layer 0.4)
            // get a multiplicative boost. This ensures date-relevant memories rise
            // above semantically similar but temporally wrong results.
            if !temporal_prefilter_ids.is_empty() {
                let mut boosted_count = 0;
                for id in &temporal_prefilter_ids {
                    if let Some(score) = fused.get_mut(id) {
                        *score *= 1.0 + crate::constants::TEMPORAL_PREFILTER_BOOST;
                        boosted_count += 1;
                    }
                    // Don't insert memories absent from fusion — date range alone
                    // without semantic/graph support is insufficient evidence
                }
                if boosted_count > 0 {
                    tracing::info!(
                        "Layer 4.45: Boosted {} memories from temporal pre-filter",
                        boosted_count
                    );
                }
            }

            // ===========================================================================
            // LAYER 4.55: TEMPORAL FACT BOOST
            // ===========================================================================
            // Source memories of matching temporal facts get a moderate boost.
            // This ensures "When did Melanie paint a sunrise?" boosts the memory that
            // recorded the event, not just memories with temporal_refs.
            if !temporal_fact_boost_ids.is_empty() {
                let mut boosted_count = 0;
                for id in &temporal_fact_boost_ids {
                    if let Some(score) = fused.get_mut(id) {
                        // Multiplicative: temporal fact match amplifies base score
                        *score *= 1.0 + crate::constants::TEMPORAL_FACT_BOOST;
                        boosted_count += 1;
                    }
                    // Don't insert absent memories — temporal fact match alone is insufficient
                }
                if boosted_count > 0 {
                    tracing::info!(
                        "Layer 4.55: Boosted {} memories from temporal fact matches",
                        boosted_count
                    );
                }
            }

            // ===========================================================================
            // LAYER 4.6: INTERFERENCE-AWARE SCORING (PIPE-3)
            // ===========================================================================
            // Research basis: Anderson & Neely (1996) - Retrieval-induced forgetting
            //
            // Retrieval is a competitive process:
            // - Memories that frequently "lose" competitions → harder to retrieve (suppress)
            // - Memories that survive despite competition → stronger/reliable (boost)
            //
            // The adjustment is based on interference history + current activation:
            // - High interference + high activation = "survivor" → boost (1.0-1.5x)
            // - High interference + low activation = "chronic loser" → suppress (0.5-1.0x)
            // - No interference history → neutral (1.0x)
            {
                let detector = self.interference_detector.read();

                // Compute max score once for normalization
                let max_score = fused
                    .values()
                    .copied()
                    .fold(0.0_f32, |a, b| a.max(b))
                    .max(0.01);

                // Collect adjustments first to avoid borrow issues
                let adjustments: Vec<_> = fused
                    .iter()
                    .map(|(id, &score)| {
                        let current_activation = (score / max_score).clamp(0.0, 1.0);
                        let adjustment = detector
                            .calculate_retrieval_adjustment(&id.0.to_string(), current_activation);
                        (id.clone(), adjustment)
                    })
                    .filter(|(_, adj)| (*adj - 1.0).abs() > 0.01)
                    .collect();

                // Apply adjustments
                let adjusted_count = adjustments.len();
                for (id, adjustment) in adjustments {
                    if let Some(score) = fused.get_mut(&id) {
                        *score *= adjustment;
                    }
                }

                if adjusted_count > 0 {
                    tracing::debug!(
                        "Layer 4.6 (PIPE-3): Applied interference adjustments to {} memories",
                        adjusted_count
                    );
                }
            }

            // ===========================================================================
            // LAYER 4.7: PROSPECTIVE SIGNAL BOOST ("Future Informs Present")
            // ===========================================================================
            // Research basis: Einstein & McDaniel (2005) - Prospective Memory
            //
            // Active goals and pending intentions shape what we remember.
            // When context-triggered prospective tasks match the current query,
            // memories related to those intentions become more accessible —
            // just as prospective memory primes retrospective recall in humans.
            //
            // Signals come from ProspectiveTasks that matched the current query
            // via keyword or semantic similarity (built in recall handler C5).
            if let Some(ref signals) = query.prospective_signals {
                if !signals.is_empty() {
                    use crate::constants::{PROSPECTIVE_BOOST_MAX, PROSPECTIVE_BOOST_PER_MATCH};

                    // Tokenize all signals into unique terms (skip noise words < 3 chars)
                    let signal_terms: std::collections::HashSet<String> = signals
                        .iter()
                        .flat_map(|s| {
                            s.to_lowercase()
                                .split_whitespace()
                                .filter(|w| w.len() >= 3)
                                .map(|w| w.to_string())
                                .collect::<Vec<_>>()
                        })
                        .collect();

                    if !signal_terms.is_empty() {
                        let mut boosted_count = 0;
                        let ids: Vec<MemoryId> = fused.keys().cloned().collect();

                        for id in &ids {
                            if let Some(content) = get_content(id) {
                                let content_lower = content.to_lowercase();
                                let content_words = tokenize_words(&content_lower);
                                let match_count = signal_terms
                                    .iter()
                                    .filter(|term| content_words.contains(term.as_str()))
                                    .count();

                                if match_count > 0 {
                                    // Sqrt scaling: diminishing returns for additional matches
                                    let boost_factor = 1.0
                                        + (PROSPECTIVE_BOOST_PER_MATCH
                                            * (match_count as f32).sqrt())
                                        .min(PROSPECTIVE_BOOST_MAX);
                                    *fused.get_mut(id).unwrap() *= boost_factor;
                                    boosted_count += 1;
                                }
                            }
                        }

                        if boosted_count > 0 {
                            tracing::info!(
                                "Layer 4.7: Boosted {} memories from {} prospective signal terms",
                                boosted_count,
                                signal_terms.len()
                            );
                        }
                    }
                }
            }

            // ===========================================================================
            // LAYER 4.8: SEMANTIC FACT SOURCE BOOST
            // ===========================================================================
            // Consolidated facts represent stable knowledge. When query entities match
            // fact entities, the source memories that generated those facts contain the
            // richest context — they should rank higher.
            //
            // Conservative: only boosts memories already in fused set (does NOT inject
            // new candidates). Facts validate existing retrieval signals, not override.
            if !fact_source_boosts.is_empty() {
                let mut boosted_count = 0;
                for (id, boost) in &fact_source_boosts {
                    if let Some(score) = fused.get_mut(id) {
                        // Multiplicative: fact source validation amplifies base score
                        *score *= 1.0 + boost;
                        boosted_count += 1;
                    }
                }
                if boosted_count > 0 {
                    tracing::info!(
                        "Layer 4.8: Boosted {} memories from semantic fact sources",
                        boosted_count
                    );
                }
            }

            // ===========================================================================
            // LAYER 4.9: ONTOLOGICAL RE-RANKING
            // ===========================================================================
            // Boost memories connected to type-matching entities. Multiplicative boost
            // on fused score. Only active when ontological intent has sufficient confidence.
            //
            // Reference: Collins & Quillian (1969) — type-plausible paths retrieved faster
            // Pre-sort and limit candidates before expensive re-ranking.
            // Only look up graph entities for the top 2x max_results candidates,
            // not all fused results (avoids 100s of RocksDB reads).
            let mut res: Vec<_> = fused.into_iter().collect();
            res.sort_by(|a, b| b.1.total_cmp(&a.1));
            let rerank_budget = query.max_results * 2;

            if use_ontology_rerank && !onto_intent.expected_labels.is_empty() {
                if let Some(graph) = self.graph_memory.as_ref() {
                    let g = graph.read();
                    let mut boosted_count = 0usize;
                    for (_mem_id, fused_score) in res.iter_mut().take(rerank_budget) {
                        if let Ok(Some(episode)) = g.get_episode(&_mem_id.0) {
                            let type_matches = episode
                                .entity_refs
                                .iter()
                                .filter(|uuid| {
                                    g.get_entity(uuid)
                                        .ok()
                                        .flatten()
                                        .map(|e| {
                                            e.labels.iter().any(|l| {
                                                onto_intent
                                                    .expected_labels
                                                    .iter()
                                                    .any(|exp| l.matches_with_hierarchy(exp))
                                            })
                                        })
                                        .unwrap_or(false)
                                })
                                .count();
                            if type_matches > 0 {
                                let boost = (type_matches as f32
                                    * crate::constants::ONTOLOGICAL_RERANK_BOOST)
                                    .min(crate::constants::ONTOLOGICAL_RERANK_MAX);
                                *fused_score *= 1.0 + boost;
                                boosted_count += 1;
                                crate::metrics::ONTOLOGICAL_RERANK_BOOST_APPLIED
                                    .observe(boost as f64);
                            }
                        }
                    }
                    if boosted_count > 0 {
                        tracing::debug!(
                            "Layer 4.9: Ontological re-rank boosted {} of {} candidates (labels={:?})",
                            boosted_count,
                            rerank_budget.min(res.len()),
                            onto_intent.expected_labels
                        );
                    }
                    // Re-sort after boosting since ranks may have changed
                    res.sort_by(|a, b| b.1.total_cmp(&a.1));
                }
            }

            res.truncate(query.max_results);
            tracing::debug!("Layer 4: {} fused results", res.len());
            (res, heb)
        };

        let t_fusion = recall_start.elapsed();
        tracing::info!(
            fusion_ms = format!("{:.2}", (t_fusion - t_vector).as_secs_f64() * 1000.0),
            cumulative_ms = format!("{:.2}", t_fusion.as_secs_f64() * 1000.0),
            fused_results = memory_ids.len(),
            "recall [layer:4] BM25 + RRF fusion + boosts + interference"
        );

        // Fetch memories with cache-aware strategy
        // CRITICAL: Apply filters after fetching to ensure mission_id, robot_id etc. are respected
        let mut memories = Vec::new();
        let mut sources = Vec::new();
        let mut cache_hits = 0;
        let mut storage_fetches = 0;
        let mut filtered_out = 0;

        // Layer 5: Unified scoring with hebbian + recency + emotional + feedback signals
        // All signals are multiplicative on the base score to preserve RRF ranking.
        // Formula: base × importance × (1 + recency + arousal + credibility + temporal) × feedback
        let now = chrono::Utc::now();

        // PIPE-9: Get feedback store guard for momentum-based scoring
        // Acquire once outside the loop to avoid repeated locking
        let feedback_guard = self.feedback_store.as_ref().map(|fs| fs.read());

        for (memory_id, score) in memory_ids {
            // Hebbian boost from learned graph weights (multiplicative)
            let hebbian_boost = hebbian_scores.get(&memory_id).copied().unwrap_or(0.0);
            let base_score =
                score * (1.0 + hebbian_boost * crate::constants::HEBBIAN_ASSOCIATION_WEIGHT);

            // Helper to apply unified scoring (all multiplicative on base score)
            let recency_scale_override = query.recency_weight;
            let with_unified_score = |mem: &SharedMemory, base: f32| -> SharedMemory {
                use crate::constants::*;

                // Recency: exponential decay, multiplicative factor
                let hours_old = (now - mem.created_at).num_hours().max(0) as f32;
                let recency_scale = recency_scale_override.unwrap_or(RECENCY_BOOST_SCALE);
                let recency_factor = (-RECENCY_DECAY_RATE * hours_old).exp() * recency_scale;

                // Emotional arousal: high arousal = more salient
                // Reference: LaBar & Cabeza (2006) — emotionally arousing events better remembered
                let arousal_factor = mem
                    .experience
                    .context
                    .as_ref()
                    .map(|c| c.emotional.arousal * AROUSAL_BOOST_SCALE)
                    .unwrap_or(0.0);

                // Source credibility: credible sources weighted higher
                let credibility_factor = mem
                    .experience
                    .context
                    .as_ref()
                    .map(|c| (c.source.credibility - 0.5).max(0.0) * CREDIBILITY_BOOST_SCALE)
                    .unwrap_or(0.0);

                // TEMPORAL MATCH (TEMPR approach for multi-hop temporal retrieval)
                //
                // Three-tier strategy:
                // 1. If memory has explicit temporal_refs → match against query refs (highest signal)
                // 2. If query is filtering BY date → use created_at proximity as fallback
                // 3. If query is seeking FOR a date (WhenQuestion) → skip boost entirely
                let temporal_factor = if has_temporal_query && !temporal_ctx.is_seeking_query {
                    let mut best_match = 0.0_f32;

                    // Tier 1: Explicit temporal_refs on the memory
                    if !mem.experience.temporal_refs.is_empty() {
                        for mem_ref in &mem.experience.temporal_refs {
                            for query_ref in &temporal_ctx.extraction.refs {
                                if mem_ref == &query_ref.date.to_string() {
                                    best_match = best_match.max(TEMPORAL_MATCH_BOOST_EXACT);
                                } else if let Ok(mem_date) =
                                    chrono::NaiveDate::parse_from_str(mem_ref, "%Y-%m-%d")
                                {
                                    let days_diff = (mem_date - query_ref.date).num_days().abs();
                                    if days_diff <= 7 {
                                        let proximity = TEMPORAL_MATCH_BOOST_WEEK
                                            * (1.0 - days_diff as f32 / 7.0);
                                        best_match = best_match.max(proximity);
                                    } else if days_diff <= 30 {
                                        let proximity = TEMPORAL_MATCH_BOOST_MONTH
                                            * (1.0 - days_diff as f32 / 30.0);
                                        best_match = best_match.max(proximity);
                                    }
                                }
                            }
                        }
                    }

                    // Tier 2: created_at proximity fallback for filtering queries
                    // Most memories don't have temporal_refs, so use created_at as proxy
                    if best_match == 0.0 && temporal_ctx.is_filtering_query {
                        if let Some((range_start, range_end)) = temporal_ctx.date_range {
                            let created_date = mem.created_at.date_naive();
                            let range_mid = range_start
                                + chrono::Duration::days((range_end - range_start).num_days() / 2);
                            let days_off = (created_date - range_mid).num_days().abs();
                            let range_half = ((range_end - range_start).num_days() / 2).max(1);

                            if days_off <= range_half {
                                // Within range: full boost
                                best_match = TEMPORAL_MATCH_BOOST_EXACT;
                            } else if days_off <= range_half + 7 {
                                // Near range: decaying week boost
                                let proximity = TEMPORAL_MATCH_BOOST_WEEK
                                    * (1.0 - (days_off - range_half) as f32 / 7.0);
                                best_match = best_match.max(proximity);
                            } else if days_off <= range_half + 30 {
                                // Further out: decaying month boost
                                let proximity = TEMPORAL_MATCH_BOOST_MONTH
                                    * (1.0 - (days_off - range_half) as f32 / 30.0);
                                best_match = best_match.max(proximity);
                            }
                        }
                    }

                    best_match
                } else {
                    0.0
                };

                // FEEDBACK MOMENTUM (PIPE-9)
                // Symmetric ±15% multiplicative adjustment
                let feedback_multiplier = if let Some(ref guard) = feedback_guard {
                    if let Some(fm) = guard.get_momentum(&mem.id) {
                        let momentum = fm.ema_with_decay();
                        if momentum < 0.0 {
                            1.0 + (momentum * FEEDBACK_MOMENTUM_SCALE).max(-FEEDBACK_MOMENTUM_SCALE)
                        } else {
                            1.0 + (momentum * FEEDBACK_MOMENTUM_SCALE).min(FEEDBACK_MOMENTUM_SCALE)
                        }
                    } else {
                        1.0
                    }
                } else {
                    1.0
                };

                // Importance: scale base score by learned importance (7 factors)
                let importance_factor =
                    SCORING_IMPORTANCE_FLOOR + mem.importance() * SCORING_IMPORTANCE_RANGE;

                // Unified multiplicative scoring:
                // base × importance × (1 + recency + arousal + credibility + temporal) × feedback
                // All boost factors are additive within the parenthetical, then multiplicative
                // on the base. This preserves RRF ranking while allowing signals to modulate.
                let combined_boost =
                    1.0 + recency_factor + arousal_factor + credibility_factor + temporal_factor;
                let final_score = base * importance_factor * combined_boost * feedback_multiplier;

                let mut cloned: Memory = mem.as_ref().clone();
                cloned.set_score(final_score);
                Arc::new(cloned)
            };

            // Try working memory first (hot cache)
            if let Some(memory) = self.working_memory.read().get(&memory_id) {
                // CRITICAL FIX: Apply filters before adding to results
                if self.retriever.matches_filters(&memory, &vector_query) {
                    memories.push(with_unified_score(&memory, base_score));
                    if !sources.contains(&"working") {
                        sources.push("working");
                    }
                    cache_hits += 1;
                } else {
                    filtered_out += 1;
                }
                continue;
            }

            // Try session memory second (warm cache)
            if let Some(memory) = self.session_memory.read().get(&memory_id) {
                // CRITICAL FIX: Apply filters before adding to results
                if self.retriever.matches_filters(&memory, &vector_query) {
                    memories.push(with_unified_score(&memory, base_score));
                    if !sources.contains(&"session") {
                        sources.push("session");
                    }
                    cache_hits += 1;
                } else {
                    filtered_out += 1;
                }
                continue;
            }

            // Cold path: Fetch from RocksDB storage (expensive deserialization)
            match self.retriever.get_from_storage(&memory_id) {
                Ok(memory) => {
                    // CRITICAL FIX: Apply filters before adding to results
                    if self.retriever.matches_filters(&memory, &vector_query) {
                        // Reuse unified scoring (includes feedback_multiplier)
                        let shared = Arc::new(memory);
                        memories.push(with_unified_score(&shared, base_score));
                        if !sources.contains(&"longterm") {
                            sources.push("longterm");
                        }
                        storage_fetches += 1;
                    } else {
                        filtered_out += 1;
                    }
                }
                Err(e) => {
                    tracing::warn!(
                        memory_id = %memory_id.0,
                        error = %e,
                        "Stale vector reference — cleaning up orphaned index entry"
                    );
                    self.retriever.remove_memory(&memory_id);
                }
            }

            if memories.len() >= query.max_results {
                break;
            }
        }

        tracing::debug!(filtered_out = filtered_out, "Filter pass completed");

        // Log cache efficiency
        tracing::debug!(
            cache_hits = cache_hits,
            storage_fetches = storage_fetches,
            hit_rate = if cache_hits + storage_fetches > 0 {
                (cache_hits as f32 / (cache_hits + storage_fetches) as f32) * 100.0
            } else {
                0.0
            },
            "Cache-aware retrieval completed"
        );

        let t_fetch = recall_start.elapsed();
        tracing::info!(
            fetch_ms = format!("{:.2}", (t_fetch - t_fusion).as_secs_f64() * 1000.0),
            cumulative_ms = format!("{:.2}", t_fetch.as_secs_f64() * 1000.0),
            memories = memories.len(),
            cache_hits,
            storage_fetches,
            filtered_out,
            "recall [layer:5] memory fetch + unified scoring"
        );

        // Quality gate: multiplicative factor based on content richness.
        // Prevents empty/trivial memories from surfacing on recency or graph boost alone.
        // Same formula as proactive_context (Berntsen elaboration quality).
        for mem in &mut memories {
            let content_len = mem.experience.content.len() as f32;
            let has_entities = !mem.experience.entities.is_empty();
            let has_context = mem.experience.context.is_some();
            let quality = (content_len / 200.0).min(1.0)
                * (1.0
                    + if has_entities { 0.1 } else { 0.0 }
                    + if has_context { 0.1 } else { 0.0 });
            let quality_factor = quality.max(crate::constants::ELABORATION_QUALITY_MIN);
            if let Some(score) = mem.score {
                let mut cloned: Memory = mem.as_ref().clone();
                cloned.set_score(score * quality_factor);
                *mem = Arc::new(cloned);
            }
        }

        // Linguistic analysis: additive boost (5% of IC weight), not a full re-sort
        if !query_analysis.focal_entities.is_empty() {
            memories.sort_by(|a, b| {
                let score_a = a.score.unwrap_or(0.0)
                    + Self::linguistic_boost(&a.experience.content, &query_analysis) * 0.05;
                let score_b = b.score.unwrap_or(0.0)
                    + Self::linguistic_boost(&b.experience.content, &query_analysis) * 0.05;
                score_b.total_cmp(&score_a)
            });
        }

        self.logger
            .read()
            .log_retrieved(query_text, memories.len(), &sources);

        // SHO-106: Apply retrieval competition between similar memories FIRST
        // When highly similar memories are retrieved, they compete for activation
        // PIPE-10: Competition must happen BEFORE coactivation - we only want to
        // strengthen associations between memories that "won" the competition.
        // Suppressed memories should not be coactivated (Hebbian "losers don't learn").
        if memories.len() >= 2 {
            // Use actual pipeline scores for competition, not position-based proxies.
            // Previously used 1.0 - (i/n)*0.3 which compressed all scores to [0.7, 1.0],
            // making suppression (ratio > 0.9) almost impossible.
            let candidates: Vec<(String, f32, f32)> = memories
                .iter()
                .map(|m| {
                    let pipeline_score = m.score.unwrap_or(0.0);
                    (m.id.0.to_string(), pipeline_score, pipeline_score)
                })
                .collect();

            let competition_result = self
                .interference_detector
                .write()
                .apply_retrieval_competition(&candidates, query_text);

            // Record competition event if any memories were suppressed
            if let Some(ref event) = competition_result.event {
                self.record_consolidation_event(event.clone());
            }

            // Re-order memories based on competition results (winners first)
            if !competition_result.suppressed.is_empty() {
                let winner_set: std::collections::HashSet<_> = competition_result
                    .winners
                    .iter()
                    .map(|(id, _)| id.clone())
                    .collect();

                // Keep only winners, maintain their relative order
                memories.retain(|m| winner_set.contains(&m.id.0.to_string()));

                tracing::debug!(
                    "Retrieval competition: {} memories suppressed",
                    competition_result.suppressed.len()
                );
            }

            // Persist interference records from retrieval competition
            {
                let detector = self.interference_detector.read();
                let affected_ids = detector.get_affected_ids_from_competition(&competition_result);
                if !affected_ids.is_empty() {
                    for (id, records) in detector.get_records_for_ids(&affected_ids) {
                        if let Err(e) = self.long_term_memory.save_interference_records(id, records)
                        {
                            tracing::debug!("Failed to persist competition interference: {e}");
                        }
                    }
                    let (total_events, _) = detector.stats();
                    if let Err(e) = self
                        .long_term_memory
                        .save_interference_event_count(total_events)
                    {
                        tracing::debug!("Failed to persist interference event count: {e}");
                    }
                }
            }
        }

        // Update access counts with instrumentation for consolidation events
        // (only for memories that survived competition)
        for memory in &memories {
            self.update_access_count_instrumented(memory, StrengtheningReason::Recalled);
        }

        // PIPE-10: Hebbian learning AFTER competition - only coactivate winners
        // When memories are retrieved together AND survive competition, they
        // form/strengthen edges in the memory graph. Suppressed memories don't
        // participate in coactivation (biological: "neurons that fire together
        // wire together" but suppressed neurons don't fire).
        if memories.len() >= 2 {
            if let Some(graph) = &self.graph_memory {
                let memory_uuids: Vec<uuid::Uuid> = memories.iter().map(|m| m.id.0).collect();
                match graph.read().record_memory_coactivation(&memory_uuids) {
                    Ok(edges_updated) if edges_updated > 0 => {
                        // Record consolidation events for coactivation visibility
                        for i in 0..memories.len().min(5) {
                            for j in (i + 1)..memories.len().min(5) {
                                self.record_consolidation_event(
                                    introspection::ConsolidationEvent::EdgeStrengthened {
                                        from_memory_id: memories[i].id.0.to_string(),
                                        to_memory_id: memories[j].id.0.to_string(),
                                        strength_before: 0.0,
                                        strength_after: 0.025,
                                        co_activations: 1,
                                        timestamp: chrono::Utc::now(),
                                    },
                                );
                            }
                        }
                    }
                    Err(e) => {
                        tracing::trace!("Coactivation recording failed (non-critical): {e}");
                    }
                    _ => {}
                }
            }
        }

        // Increment and persist retrieval counter
        if let Ok(count) = self.long_term_memory.increment_retrieval_count() {
            self.stats.write().total_retrievals = count;
        }

        // Expand with hierarchy context (parent chain + children)
        // This ensures semantic search also surfaces contextually related memories
        let mut seen_ids: HashSet<MemoryId> = memories.iter().map(|m| m.id.clone()).collect();
        self.expand_with_hierarchy(&mut memories, &mut seen_ids);

        // Re-sort by score and trim to max_results after expansion.
        // Expanded memories must compete on score, not get a free pass.
        if memories.len() > query.max_results {
            memories.sort_by(|a, b| b.score.unwrap_or(0.0).total_cmp(&a.score.unwrap_or(0.0)));
            memories.truncate(query.max_results);
        }

        let t_total = recall_start.elapsed();
        tracing::info!(
            post_ms = format!("{:.2}", (t_total - t_fetch).as_secs_f64() * 1000.0),
            total_ms = format!("{:.2}", t_total.as_secs_f64() * 1000.0),
            final_count = memories.len(),
            "recall [layer:post] linguistic + competition + coactivation + hierarchy === RECALL COMPLETE ==="
        );

        Ok(memories)
    }

    /// Get learning velocity statistics for a memory
    ///
    /// Returns information about recent learning activity for this memory,
    /// useful for debugging/introspection.
    pub fn get_learning_velocity(
        &self,
        user_id: &str,
        memory_id: &str,
        hours: i64,
    ) -> Result<learning_history::LearningVelocity> {
        self.learning_history
            .memory_learning_velocity(user_id, memory_id, hours)
    }

    /// Get learning history statistics for a user
    pub fn get_learning_stats(&self, user_id: &str) -> Result<learning_history::LearningStats> {
        self.learning_history.stats(user_id)
    }

    /// Get recent learning events for a user
    pub fn get_learning_events(
        &self,
        user_id: &str,
        since: chrono::DateTime<chrono::Utc>,
        limit: usize,
    ) -> Result<Vec<learning_history::StoredLearningEvent>> {
        let mut events = self.learning_history.events_since(user_id, since)?;
        events.truncate(limit);
        Ok(events)
    }

    // ==========================================================================
    // TEMPORAL FACT EXTRACTION (for multi-hop temporal queries)
    // ==========================================================================

    /// Extract and store temporal facts from a memory
    ///
    /// Call this after remember() when you have access to user_id.
    /// Extracts facts like "Melanie is planning camping next month" and stores them
    /// with resolved absolute dates for accurate multi-hop retrieval.
    pub fn store_temporal_facts_for_memory(
        &self,
        user_id: &str,
        memory_id: &MemoryId,
        content: &str,
        entities: &[String],
        created_at: chrono::DateTime<chrono::Utc>,
    ) -> Result<usize> {
        let facts =
            temporal_facts::extract_temporal_facts(content, memory_id, created_at, entities);
        if facts.is_empty() {
            return Ok(0);
        }

        let stored = self.temporal_fact_store.store_batch(user_id, &facts)?;
        if stored > 0 {
            tracing::debug!(
                user_id = user_id,
                memory_id = %memory_id.0,
                facts_stored = stored,
                "Stored temporal facts for memory"
            );
        }
        Ok(stored)
    }

    /// Find temporal facts by entity and event keywords
    ///
    /// Used for multi-hop queries like "When did Melanie paint a sunrise?"
    /// Returns facts sorted by conversation date (earliest first for planning queries).
    pub fn find_temporal_facts(
        &self,
        user_id: &str,
        entity: &str,
        event_keywords: &[&str],
        event_type: Option<temporal_facts::EventType>,
    ) -> Result<Vec<temporal_facts::TemporalFact>> {
        self.temporal_fact_store.find_by_entity_and_event(
            user_id,
            entity,
            event_keywords,
            event_type,
        )
    }

    /// List all temporal facts for a user
    pub fn list_temporal_facts(
        &self,
        user_id: &str,
        limit: usize,
    ) -> Result<Vec<temporal_facts::TemporalFact>> {
        self.temporal_fact_store.list(user_id, limit)
    }

    /// Calculate linguistic boost based on focal entity matches
    fn linguistic_boost(content: &str, analysis: &query_parser::QueryAnalysis) -> f32 {
        let content_lower = content.to_lowercase();
        let mut boost = 0.0;

        for entity in &analysis.focal_entities {
            if content_lower.contains(&entity.text) {
                boost += entity.ic_weight;
            }
        }

        for modifier in &analysis.discriminative_modifiers {
            if content_lower.contains(&modifier.text) {
                boost += 1.7; // IC_ADJECTIVE
            }
        }

        boost
    }

    /// Compute SHA256 hash of text for stable cache keys
    ///
    /// Unlike std::hash::DefaultHasher, SHA256 produces deterministic hashes
    /// across process restarts and Rust versions. This is critical for:
    /// - Embedding cache persistence (future feature)
    /// - Consistent behavior across restarts
    /// - Avoiding cache key collisions
    #[inline]
    fn sha256_hash(text: &str) -> [u8; 32] {
        let mut hasher = Sha256::new();
        hasher.update(text.as_bytes());
        hasher.finalize().into()
    }

    /// Forget memories based on criteria
    /// Thread-safe: uses interior mutability for all internal state
    pub fn forget(&self, criteria: ForgetCriteria) -> Result<usize> {
        let forgotten_count = match criteria {
            ForgetCriteria::ById(memory_id) => {
                // Delete a single memory by ID from all tiers, tracking which tiers had it
                let mut deleted_from_any = false;
                let mut was_in_working = false;
                let mut was_in_session = false;
                let mut was_in_longterm = false;

                // Remove from working memory
                if self.working_memory.write().remove(&memory_id).is_ok() {
                    deleted_from_any = true;
                    was_in_working = true;
                }

                // Remove from session memory
                if self.session_memory.write().remove(&memory_id).is_ok() {
                    deleted_from_any = true;
                    was_in_session = true;
                }

                // Remove from long-term storage
                if self.long_term_memory.delete(&memory_id).is_ok() {
                    deleted_from_any = true;
                    was_in_longterm = true;
                }

                // Remove from vector index (soft delete) - CRITICAL for semantic search
                // This marks the vector as deleted so it won't appear in search results
                let was_indexed = self.retriever.remove_memory(&memory_id);

                // Clean up knowledge graph episode and sourced edges
                if let Some(graph) = &self.graph_memory {
                    if let Err(e) = graph.read().delete_episode(&memory_id.0) {
                        tracing::warn!(
                            memory_id = %memory_id.0,
                            error = %e,
                            "Failed to clean up graph episode for deleted memory"
                        );
                    }
                }

                // Clean up BM25 keyword index
                if let Err(e) = self.hybrid_search.remove_memory(&memory_id) {
                    tracing::warn!(
                        memory_id = %memory_id.0,
                        error = %e,
                        "Failed to clean BM25 index for deleted memory"
                    );
                }

                // Clean up interference records
                self.cleanup_interference_for_ids(&[memory_id.clone()]);

                // Update stats - decrement each tier count that had this memory
                if deleted_from_any {
                    let mut stats = self.stats.write();
                    stats.total_memories = stats.total_memories.saturating_sub(1);
                    if was_in_working {
                        stats.working_memory_count = stats.working_memory_count.saturating_sub(1);
                    }
                    if was_in_session {
                        stats.session_memory_count = stats.session_memory_count.saturating_sub(1);
                    }
                    if was_in_longterm {
                        stats.long_term_memory_count =
                            stats.long_term_memory_count.saturating_sub(1);
                    }
                    if was_indexed {
                        stats.vector_index_count = stats.vector_index_count.saturating_sub(1);
                    }
                    1
                } else {
                    0
                }
            }
            ForgetCriteria::OlderThan(days) => {
                let cutoff = chrono::Utc::now() - chrono::Duration::days(days as i64);

                // Remove from working memory
                let working_removed = self.working_memory.write().remove_older_than(cutoff)?;

                // Remove from session memory
                let session_removed = self.session_memory.write().remove_older_than(cutoff)?;

                // Mark as forgotten in long-term (don't delete, just flag)
                let flagged_ids = self.long_term_memory.mark_forgotten_by_age(cutoff)?;
                let lt_flagged = flagged_ids.len();

                // Clean up secondary indices for soft-forgotten memories
                for id in &flagged_ids {
                    self.retriever.remove_memory(id);
                    let _ = self.hybrid_search.remove_memory(id);
                }
                self.cleanup_graph_for_ids(&flagged_ids);
                self.cleanup_interference_for_ids(&flagged_ids);

                // Update stats for hard-deleted and soft-deleted tiers
                {
                    let removed = working_removed + session_removed + lt_flagged;
                    if removed > 0 {
                        let mut stats = self.stats.write();
                        stats.working_memory_count =
                            stats.working_memory_count.saturating_sub(working_removed);
                        stats.session_memory_count =
                            stats.session_memory_count.saturating_sub(session_removed);
                        stats.long_term_memory_count =
                            stats.long_term_memory_count.saturating_sub(lt_flagged);
                        stats.total_memories = stats.total_memories.saturating_sub(removed);
                    }
                }

                lt_flagged
            }
            ForgetCriteria::LowImportance(threshold) => {
                let working_removed = self
                    .working_memory
                    .write()
                    .remove_below_importance(threshold)?;
                let session_removed = self
                    .session_memory
                    .write()
                    .remove_below_importance(threshold)?;
                let flagged_ids = self
                    .long_term_memory
                    .mark_forgotten_by_importance(threshold)?;
                let lt_flagged = flagged_ids.len();

                // Clean up secondary indices for soft-forgotten memories
                for id in &flagged_ids {
                    self.retriever.remove_memory(id);
                    let _ = self.hybrid_search.remove_memory(id);
                }
                self.cleanup_graph_for_ids(&flagged_ids);
                self.cleanup_interference_for_ids(&flagged_ids);

                // Update stats for hard-deleted and soft-deleted tiers
                {
                    let removed = working_removed + session_removed + lt_flagged;
                    if removed > 0 {
                        let mut stats = self.stats.write();
                        stats.working_memory_count =
                            stats.working_memory_count.saturating_sub(working_removed);
                        stats.session_memory_count =
                            stats.session_memory_count.saturating_sub(session_removed);
                        stats.long_term_memory_count =
                            stats.long_term_memory_count.saturating_sub(lt_flagged);
                        stats.total_memories = stats.total_memories.saturating_sub(removed);
                    }
                }

                lt_flagged
            }
            ForgetCriteria::Pattern(pattern) => {
                // Remove memories matching pattern
                self.forget_by_pattern(&pattern)?
            }
            ForgetCriteria::ByTags(tags) => {
                // Remove memories matching ANY of the specified tags
                self.forget_by_tags(&tags)?
            }
            ForgetCriteria::ByDateRange { start, end } => {
                // Remove memories within the date range
                self.forget_by_date_range(start, end)?
            }
            ForgetCriteria::ByType(exp_type) => {
                // Remove memories of a specific type
                self.forget_by_type(exp_type)?
            }
            ForgetCriteria::All => {
                // GDPR: Clear ALL memories for the user
                self.forget_all()?
            }
        };

        // Commit BM25 changes after any deletion to make removals visible
        if forgotten_count > 0 {
            if let Err(e) = self.hybrid_search.commit_and_reload() {
                tracing::warn!(error = %e, "Failed to commit BM25 after forget");
            }
        }

        Ok(forgotten_count)
    }

    /// Get memory statistics
    ///
    /// Returns current stats with fresh average_importance calculated from storage.
    /// Most counters are cached in-memory for performance, but importance is
    /// recalculated to ensure accuracy after memory modifications.
    pub fn stats(&self) -> MemoryStats {
        let mut stats = self.stats.read().clone();

        // Recalculate average_importance from storage for accuracy
        // This ensures importance reflects current memory state after adds/deletes
        if let Ok(storage_stats) = self.long_term_memory.get_stats() {
            stats.average_importance = storage_stats.average_importance;
        }

        stats
    }

    /// Export visualization graph as DOT format for Graphviz
    pub fn export_visualization_dot(&self) -> String {
        self.logger.read().graph.to_dot()
    }

    /// Build visualization graph from current memory state
    /// Call this to populate the visualization graph with all current memories
    pub fn build_visualization_graph(&self) -> Result<visualization::GraphStats> {
        let mut logger = self.logger.write();

        // Add working memory entries directly to the graph (bypasses enabled check)
        for memory in self.working_memory.read().all_memories() {
            logger.graph.add_memory(&memory, "working");
        }

        // Add session memory entries
        for memory in self.session_memory.read().all_memories() {
            logger.graph.add_memory(&memory, "session");
        }

        // Add long-term memory entries
        for memory in self.long_term_memory.get_all()? {
            logger.graph.add_memory(&memory, "longterm");
        }

        Ok(logger.get_stats())
    }

    /// Get reference to embedder for graph-aware retrieval
    pub fn get_embedder(&self) -> &dyn Embedder {
        self.embedder.as_ref()
    }

    /// Compute embedding for arbitrary text (for external use like prospective memory)
    pub fn compute_embedding(&self, text: &str) -> Result<Vec<f32>> {
        self.embedder.encode(text)
    }

    /// Get all memories across all tiers for graph-aware retrieval
    /// Deduplicates by memory ID, preferring working > session > long-term
    pub fn get_all_memories(&self) -> Result<Vec<SharedMemory>> {
        use std::collections::HashSet;
        let mut seen_ids: HashSet<MemoryId> = HashSet::new();
        let mut all_memories = Vec::new();

        // Collect from working memory (highest priority - most recent/active)
        {
            let working = self.working_memory.read();
            for mem in working.all_memories() {
                if seen_ids.insert(mem.id.clone()) {
                    all_memories.push(mem);
                }
            }
        }

        // Collect from session memory (medium priority)
        {
            let session = self.session_memory.read();
            for mem in session.all_memories() {
                if seen_ids.insert(mem.id.clone()) {
                    all_memories.push(mem);
                }
            }
        }

        // Collect from long-term memory (lowest priority - wrap in Arc)
        {
            let longterm_mems = self.long_term_memory.get_all()?;
            for mem in longterm_mems {
                if seen_ids.insert(mem.id.clone()) {
                    all_memories.push(Arc::new(mem));
                }
            }
        }

        Ok(all_memories)
    }

    /// Find a memory by UUID prefix across all tiers.
    ///
    /// Accepts both full UUIDs and 8+ char hex prefixes (as displayed by MCP tools).
    /// Searches working → session → long-term memory with deduplication.
    /// Returns `Err` for ambiguous prefixes (multiple matches).
    pub fn find_memory_by_prefix(&self, id_prefix: &str) -> Result<Option<SharedMemory>> {
        // Fast path: try full UUID first via direct lookup
        if let Ok(uuid) = uuid::Uuid::parse_str(id_prefix) {
            let target_id = MemoryId(uuid);
            let all = self.get_all_memories()?;
            return Ok(all.into_iter().find(|m| m.id == target_id));
        }

        // Prefix search across all tiers
        let prefix_lower = id_prefix.to_lowercase();
        let all_memories = self.get_all_memories()?;
        let matches: Vec<SharedMemory> = all_memories
            .into_iter()
            .filter(|m| {
                m.id.0
                    .to_string()
                    .replace('-', "")
                    .to_lowercase()
                    .starts_with(&prefix_lower)
            })
            .collect();

        match matches.len() {
            0 => Ok(None),
            1 => Ok(Some(matches.into_iter().next().unwrap())),
            n => Err(anyhow::anyhow!(
                "Ambiguous memory ID prefix '{}': matches {} memories",
                id_prefix,
                n
            )),
        }
    }

    /// Get memories from working memory tier (highest activation, most recent)
    pub fn get_working_memories(&self) -> Vec<SharedMemory> {
        let working = self.working_memory.read();
        working.all_memories()
    }

    /// Get memories from session memory tier (medium-term, consolidated)
    pub fn get_session_memories(&self) -> Vec<SharedMemory> {
        let session = self.session_memory.read();
        session.all_memories()
    }

    /// Get memories from long-term memory tier (persistent, lower activation)
    /// Returns up to `limit` memories to avoid overwhelming responses
    pub fn get_longterm_memories(&self, limit: usize) -> Result<Vec<Memory>> {
        let all = self.long_term_memory.get_all()?;
        Ok(all.into_iter().take(limit).collect())
    }

    /// Calculate temporal relevance based on memory age (ENTERPRISE FEATURE)
    ///
    /// Implements exponential decay curve for time-aware memory retrieval:
    /// - 0-7 days: Full relevance (1.0) - recent memories
    /// - 8-30 days: High relevance (0.7) - medium-term memories
    /// - 31-90 days: Moderate relevance (0.4) - older memories
    /// - 90+ days: Low relevance (0.2) - ancient memories
    ///
    /// This ensures recent experiences are prioritized while maintaining
    /// access to historical context when needed.
    fn calculate_temporal_relevance(age_days: i64) -> f32 {
        match age_days {
            0..=7 => 1.0,   // Recent: Full weight
            8..=30 => 0.7,  // Medium-term: 70% weight
            31..=90 => 0.4, // Old: 40% weight
            _ => 0.2,       // Ancient: 20% weight (never completely forgotten)
        }
    }

    /// Expand retrieved memories with their hierarchy context
    ///
    /// When a memory is retrieved, its parent chain and children are also
    /// contextually relevant. This method adds them to the result set with
    /// slightly boosted importance (hierarchy context is valuable).
    ///
    /// Hierarchy expansion depth is limited to prevent explosion:
    /// - Parents: Full chain up to root (usually shallow)
    /// - Children: Direct children only (1 level)
    fn expand_with_hierarchy(
        &self,
        memories: &mut Vec<SharedMemory>,
        seen_ids: &mut HashSet<MemoryId>,
    ) {
        // Skip if no memories to expand
        if memories.is_empty() {
            return;
        }

        // Collect IDs to expand (copy to avoid borrow issues)
        let ids_to_expand: Vec<MemoryId> = memories.iter().map(|m| m.id.clone()).collect();

        // Expand each memory with its hierarchy
        for memory_id in ids_to_expand {
            // Get parent chain
            if let Ok(ancestors) = self.long_term_memory.get_ancestors(&memory_id) {
                for ancestor in ancestors {
                    if seen_ids.insert(ancestor.id.clone()) {
                        // Boost ancestor importance slightly (context is valuable)
                        let new_importance = (ancestor.importance() * 1.1).min(1.0);
                        let mut shared = Arc::new(ancestor);
                        Arc::make_mut(&mut shared).set_importance(new_importance);
                        memories.push(shared);
                    }
                }
            }

            // Get direct children
            if let Ok(children) = self.long_term_memory.get_children(&memory_id) {
                for child in children {
                    if seen_ids.insert(child.id.clone()) {
                        // Boost child importance slightly
                        let new_importance = (child.importance() * 1.05).min(1.0);
                        let mut shared = Arc::new(child);
                        Arc::make_mut(&mut shared).set_importance(new_importance);
                        memories.push(shared);
                    }
                }
            }
        }
    }

    /// Calculate importance of an experience using multi-factor analysis
    fn calculate_importance(&self, experience: &Experience) -> f32 {
        let mut factors = Vec::new();

        // Factor 1: Experience type base score (0.0 - 0.3)
        let type_score = match experience.experience_type {
            ExperienceType::Decision => 0.3,
            ExperienceType::Error => 0.25,
            ExperienceType::Learning => 0.25,
            ExperienceType::Discovery => 0.2,
            ExperienceType::Pattern => 0.2,
            ExperienceType::Task => 0.15,
            ExperienceType::Conversation => 0.1,
            ExperienceType::Context => 0.1,
            _ => 0.05,
        };
        factors.push(("type", type_score));

        // Factor 2: Content richness (0.0 - 0.25)
        let _content_length = experience.content.len();
        let word_count = experience.content.split_whitespace().count();
        let richness_score = if word_count > 50 {
            0.25
        } else if word_count > 20 {
            0.15
        } else if word_count > 5 {
            0.08
        } else {
            0.02
        };
        factors.push(("richness", richness_score));

        // Factor 3: Entity density (0.0 - 0.2)
        let entity_score = if experience.entities.len() > 10 {
            0.2
        } else if experience.entities.len() > 5 {
            0.15
        } else if experience.entities.len() > 2 {
            0.1
        } else if !experience.entities.is_empty() {
            0.05
        } else {
            0.0
        };
        factors.push(("entities", entity_score));

        // Factor 4: Context depth (0.0 - 0.2)
        let context_score = if let Some(ctx) = &experience.context {
            let mut score: f32 = 0.0;

            // Rich semantic context
            if !ctx.semantic.concepts.is_empty() {
                score += 0.05;
            }
            if !ctx.semantic.tags.is_empty() {
                score += 0.03;
            }
            if !ctx.semantic.related_concepts.is_empty() {
                score += 0.04;
            }

            // Project/workspace context
            if ctx.project.project_id.is_some() {
                score += 0.03;
            }

            // Code context
            if ctx.code.current_file.is_some() {
                score += 0.03;
            }

            // Document citations
            if !ctx.document.citations.is_empty() {
                score += 0.02;
            }

            score.min(0.2)
        } else {
            0.0
        };
        factors.push(("context", context_score));

        // Factor 5: Metadata signals (0.0 - 0.15)
        let mut metadata_score: f32 = 0.0;

        if experience.metadata.contains_key("priority") {
            if let Some(priority) = experience.metadata.get("priority") {
                metadata_score += match priority.as_str() {
                    "critical" => 0.15,
                    "high" => 0.10,
                    "medium" => 0.05,
                    _ => 0.0,
                };
            }
        }

        if experience.metadata.contains_key("unexpected") {
            metadata_score += 0.08;
        }

        if experience.metadata.contains_key("breakthrough") {
            metadata_score += 0.12;
        }

        if experience.metadata.get("role") == Some(&"user".to_string()) {
            metadata_score += 0.02; // User messages slightly more important
        }

        factors.push(("metadata", metadata_score.min(0.15)));

        // Factor 6: Embeddings quality (0.0 - 0.1)
        let embedding_score = if let Some(emb) = &experience.embeddings {
            if emb.len() >= 384 {
                // Full embedding vector
                0.1
            } else {
                0.05
            }
        } else {
            0.0
        };
        factors.push(("embeddings", embedding_score));

        // Factor 7: Content quality indicators (0.0 - 0.1)
        let content_lower = experience.content.to_lowercase();
        let mut quality_score: f32 = 0.0;

        // Technical terms indicate higher quality
        let technical_terms = [
            "algorithm",
            "architecture",
            "implementation",
            "optimization",
            "performance",
            "security",
            "database",
            "api",
            "framework",
        ];
        for term in &technical_terms {
            if content_lower.contains(term) {
                quality_score += 0.015;
            }
        }

        // Questions indicate learning/discovery
        if content_lower.contains('?') {
            quality_score += 0.02;
        }

        // Code snippets indicate actionable content
        if experience.content.contains("```")
            || experience.content.contains("fn ")
            || experience.content.contains("function ")
            || experience.content.contains("class ")
        {
            quality_score += 0.03;
        }

        factors.push(("quality", quality_score.min(0.1)));

        // Aggregate all factors
        let importance: f32 = factors.iter().map(|(_, score)| score).sum();

        // Ensure importance is in valid range [0.0, 1.0]
        let importance = importance.clamp(0.0, 1.0);

        // Log importance calculation for transparency
        if importance > 0.7 {
            debug!("High importance memory: {:.2} (type={:.2}, richness={:.2}, entities={:.2}, context={:.2})",
                importance,
                factors.iter().find(|(k, _)| *k == "type").map(|(_, v)| v).unwrap_or(&0.0),
                factors.iter().find(|(k, _)| *k == "richness").map(|(_, v)| v).unwrap_or(&0.0),
                factors.iter().find(|(k, _)| *k == "entities").map(|(_, v)| v).unwrap_or(&0.0),
                factors.iter().find(|(k, _)| *k == "context").map(|(_, v)| v).unwrap_or(&0.0)
            );
        }

        importance
    }

    /// Consolidate memories based on Cowan's model (importance + time, not size)
    ///
    /// Tier promotion criteria:
    /// - Working → Session: importance >= 0.4 AND age >= 5 minutes
    /// - Session → LongTerm: importance >= 0.6 AND age >= 1 hour
    fn consolidate_if_needed(&self) -> Result<()> {
        // Promote eligible memories from working to session (importance + time based)
        self.promote_working_to_session()?;

        // Promote eligible memories from session to long-term (importance + time based)
        self.promote_session_to_longterm()?;

        // Compress old memories if auto-compress is enabled
        if self.config.auto_compress {
            self.compress_old_memories()?;
        }

        Ok(())
    }

    /// Move memories from working to session memory (Cowan's model)
    ///
    /// Promotion criteria: importance >= TIER_PROMOTION_WORKING_IMPORTANCE
    /// AND age >= TIER_PROMOTION_WORKING_AGE_SECS
    fn promote_working_to_session(&self) -> Result<()> {
        let now = chrono::Utc::now();
        let min_age = chrono::Duration::seconds(TIER_PROMOTION_WORKING_AGE_SECS);

        // Find eligible memories (importance + time threshold, with graph-adjusted threshold)
        let to_promote: Vec<SharedMemory> = {
            let working = self.working_memory.read();
            working
                .all_memories()
                .into_iter()
                .filter(|m| {
                    let age = now - m.created_at;
                    let importance = m.importance();
                    let threshold =
                        self.graph_adjusted_threshold(m, TIER_PROMOTION_WORKING_IMPORTANCE);
                    importance >= threshold && age >= min_age
                })
                .collect()
        };

        if to_promote.is_empty() {
            return Ok(());
        }

        let count = to_promote.len();
        let mut working = self.working_memory.write();
        let mut session = self.session_memory.write();

        for memory in &to_promote {
            // Log promotion
            self.logger
                .write()
                .log_promoted(&memory.id, "working", "session", count);

            // Clone out of Arc and update tier before session storage
            let mut promoted_memory = (**memory).clone();
            promoted_memory.promote(); // Working -> Session
            session.add(promoted_memory)?;
            working.remove(&memory.id)?;
        }

        if count > 0 {
            let mut stats = self.stats.write();
            stats.promotions_to_session += count;
            stats.working_memory_count = stats.working_memory_count.saturating_sub(count);
            stats.session_memory_count += count;
            tracing::debug!(
                "Promoted {} memories from working to session (importance >= {}, age >= {}s)",
                count,
                TIER_PROMOTION_WORKING_IMPORTANCE,
                TIER_PROMOTION_WORKING_AGE_SECS
            );
        }
        Ok(())
    }

    /// Move memories from session to long-term storage (Cowan's model)
    ///
    /// Promotion criteria: importance >= TIER_PROMOTION_SESSION_IMPORTANCE
    /// AND age >= TIER_PROMOTION_SESSION_AGE_SECS
    fn promote_session_to_longterm(&self) -> Result<()> {
        let now = chrono::Utc::now();
        let min_age = chrono::Duration::seconds(TIER_PROMOTION_SESSION_AGE_SECS);

        // Find eligible memories (importance + time threshold, with graph-adjusted threshold)
        let to_promote: Vec<SharedMemory> = {
            let session = self.session_memory.read();
            session
                .all_memories()
                .into_iter()
                .filter(|m| {
                    let age = now - m.created_at;
                    let importance = m.importance();
                    let threshold =
                        self.graph_adjusted_threshold(m, TIER_PROMOTION_SESSION_IMPORTANCE);
                    importance >= threshold && age >= min_age
                })
                .collect()
        };

        if to_promote.is_empty() {
            return Ok(());
        }

        let count = to_promote.len();
        let mut session = self.session_memory.write();

        for memory in &to_promote {
            // Log promotion
            self.logger
                .write()
                .log_promoted(&memory.id, "session", "longterm", count);

            // Clone out of Arc and update tier before long-term storage
            let mut owned_memory = (**memory).clone();
            owned_memory.promote(); // Session -> LongTerm

            // Compress if old enough
            let compressed_memory = if self.should_compress(&owned_memory) {
                self.compressor.compress(&owned_memory)?
            } else {
                owned_memory
            };

            // Store in long-term
            self.long_term_memory.store(&compressed_memory)?;

            // PRODUCTION: Index memory in Vamana vector DB for semantic search
            if let Err(e) = self.retriever.index_memory(&compressed_memory) {
                tracing::warn!(
                    "Failed to index memory {} in vector DB: {}",
                    compressed_memory.id.0,
                    e
                );
                // Don't fail promotion if indexing fails - memory is still stored
            }

            // Remove from session
            session.remove(&memory.id)?;
        }

        if count > 0 {
            let mut stats = self.stats.write();
            stats.promotions_to_longterm += count;
            stats.session_memory_count = stats.session_memory_count.saturating_sub(count);
            stats.long_term_memory_count += count;
            tracing::debug!(
                "Promoted {} memories from session to long-term (importance >= {}, age >= {}s)",
                count,
                TIER_PROMOTION_SESSION_IMPORTANCE,
                TIER_PROMOTION_SESSION_AGE_SECS
            );
        }
        Ok(())
    }

    // =========================================================================
    // Memory-Edge Tier Coupling Methods
    // =========================================================================

    /// Calculate graph-adjusted importance threshold for tier promotion (Direction 3).
    ///
    /// Well-connected memories (many L2+ edges) get a discount on the promotion threshold.
    /// Isolated memories (entities but no edges) get a penalty.
    /// Memories with no entities are unaffected (no graph context to evaluate).
    fn graph_adjusted_threshold(&self, memory: &Memory, base_threshold: f32) -> f32 {
        use crate::constants::*;

        let graph = match &self.graph_memory {
            Some(g) => g,
            None => return base_threshold,
        };

        if memory.entity_refs.is_empty() {
            return base_threshold;
        }

        let graph_guard = graph.read();
        let mut l2_plus_count = 0usize;

        for entity_ref in &memory.entity_refs {
            if let Ok(edges) = graph_guard.get_entity_relationships(&entity_ref.entity_id) {
                for edge in &edges {
                    if matches!(
                        edge.tier,
                        crate::graph_memory::EdgeTier::L2Episodic
                            | crate::graph_memory::EdgeTier::L3Semantic
                    ) {
                        l2_plus_count += 1;
                    }
                }
            }
        }

        if l2_plus_count == 0 {
            // Memory has entities but no strong edges — penalize
            base_threshold * (1.0 + GRAPH_HEALTH_NO_EDGES_PENALTY as f32)
        } else {
            // Discount proportional to edge count, capped at saturation
            let ratio = (l2_plus_count as f64 / GRAPH_HEALTH_EDGE_SATURATION).min(1.0);
            base_threshold * (1.0 - (GRAPH_HEALTH_PROMOTION_DISCOUNT * ratio) as f32)
        }
    }

    /// Apply importance boosts to memories whose edges were promoted (Direction 1).
    ///
    /// When an edge promotes from L1→L2 or L2→L3, the memories involved get
    /// a small importance boost, reflecting that they participate in a consolidating
    /// relationship. Uses interior mutability — `set_importance` works through Arc.
    pub fn apply_edge_promotion_boosts(
        &self,
        boosts: &[crate::memory::types::EdgePromotionBoost],
    ) -> Result<usize> {
        let mut applied = 0;

        for boost in boosts {
            let memory_id = match uuid::Uuid::parse_str(&boost.memory_id) {
                Ok(uuid) => MemoryId(uuid),
                Err(_) => continue,
            };

            // Search across tiers: working → session → long-term
            let found = self
                .working_memory
                .read()
                .get(&memory_id)
                .or_else(|| self.session_memory.read().get(&memory_id));

            if let Some(memory) = found {
                let new_importance = (memory.importance() + boost.boost as f32).min(1.0);
                memory.set_importance(new_importance);
                self.record_consolidation_event(ConsolidationEvent::EdgePromotionBoostApplied {
                    memory_id: boost.memory_id.clone(),
                    entity_name: boost.entity_name.clone(),
                    old_tier: boost.old_tier.clone(),
                    new_tier: boost.new_tier.clone(),
                    importance_boost: boost.boost,
                    new_importance: new_importance as f64,
                    timestamp: chrono::Utc::now(),
                });
                applied += 1;
            } else if let Ok(memory) = self.long_term_memory.get(&memory_id) {
                let new_importance = (memory.importance() + boost.boost as f32).min(1.0);
                memory.set_importance(new_importance);
                if let Err(e) = self.long_term_memory.store(&memory) {
                    tracing::debug!(
                        "Failed to persist edge promotion boost for {}: {}",
                        boost.memory_id,
                        e
                    );
                    continue;
                }
                self.record_consolidation_event(ConsolidationEvent::EdgePromotionBoostApplied {
                    memory_id: boost.memory_id.clone(),
                    entity_name: boost.entity_name.clone(),
                    old_tier: boost.old_tier.clone(),
                    new_tier: boost.new_tier.clone(),
                    importance_boost: boost.boost,
                    new_importance: new_importance as f64,
                    timestamp: chrono::Utc::now(),
                });
                applied += 1;
            }
        }

        if applied > 0 {
            tracing::debug!(
                "Applied {} edge promotion boosts to memory importance",
                applied
            );
        }

        Ok(applied)
    }

    /// Apply compensatory boost to memories that lost all graph edges (Direction 2).
    ///
    /// When graph decay prunes edges and leaves entities orphaned, the memories
    /// referencing those entities get a small importance boost to prevent immediate
    /// decay death. This gives them one more maintenance cycle to prove value.
    pub fn compensate_orphaned_memories(&self, orphaned_entity_ids: &[String]) -> Result<usize> {
        use crate::constants::ORPHAN_COMPENSATORY_BOOST;

        if orphaned_entity_ids.is_empty() {
            return Ok(0);
        }

        let orphaned_set: std::collections::HashSet<&str> =
            orphaned_entity_ids.iter().map(|s| s.as_str()).collect();

        let mut compensated = 0;

        // Scan working + session memories for references to orphaned entities
        let tiers: Vec<Vec<SharedMemory>> = vec![
            self.working_memory.read().all_memories(),
            self.session_memory.read().all_memories(),
        ];

        for memories in &tiers {
            for memory in memories {
                let entity_count = memory
                    .entity_refs
                    .iter()
                    .filter(|e| orphaned_set.contains(e.entity_id.to_string().as_str()))
                    .count();
                if entity_count > 0 {
                    let new_importance =
                        (memory.importance() + ORPHAN_COMPENSATORY_BOOST as f32).min(1.0);
                    memory.set_importance(new_importance);
                    self.record_consolidation_event(ConsolidationEvent::GraphOrphanDetected {
                        memory_id: memory.id.0.to_string(),
                        entity_count,
                        compensatory_boost: ORPHAN_COMPENSATORY_BOOST,
                        timestamp: chrono::Utc::now(),
                    });
                    compensated += 1;
                }
            }
        }

        if compensated > 0 {
            tracing::debug!(
                "Compensated {} orphaned memories (from {} orphaned entities)",
                compensated,
                orphaned_entity_ids.len()
            );
        }

        Ok(compensated)
    }

    /// Compress old memories to save space
    fn compress_old_memories(&self) -> Result<()> {
        let cutoff =
            chrono::Utc::now() - chrono::Duration::days(self.config.compression_age_days as i64);

        // Get uncompressed old memories
        let to_compress = self.long_term_memory.get_uncompressed_older_than(cutoff)?;

        for memory in to_compress {
            let compressed = self.compressor.compress(&memory)?;
            self.long_term_memory.update(&compressed)?;
            self.stats.write().compressed_count += 1;
        }

        Ok(())
    }

    /// Check if a memory should be compressed
    fn should_compress(&self, memory: &Memory) -> bool {
        let age = chrono::Utc::now() - memory.created_at;
        age.num_days() > self.config.compression_age_days as i64 && !memory.compressed
    }

    /// Update access count for a memory (handles concurrency properly)
    /// Note: Prefer update_access_count_instrumented() for consolidation tracking
    #[allow(dead_code)]
    fn update_access_count(&self, memory_id: &MemoryId) -> Result<()> {
        // Try updating in working memory first (most common case)
        // Use write lock directly to avoid TOCTOU race condition
        {
            let mut wm = self.working_memory.write();

            if wm.contains(memory_id) {
                // Memory found in working memory - update and return
                return wm
                    .update_access(memory_id)
                    .map_err(|e| anyhow::anyhow!("Failed to update working memory access: {e}"));
            }
        } // Release write lock

        // Try session memory
        {
            let mut sm = self.session_memory.write();

            if sm.contains(memory_id) {
                return sm
                    .update_access(memory_id)
                    .map_err(|e| anyhow::anyhow!("Failed to update session memory access: {e}"));
            }
        } // Release write lock

        // Try long-term memory (has its own internal locking)
        self.long_term_memory
            .update_access(memory_id)
            .map_err(|e| anyhow::anyhow!("Failed to update long-term memory access: {e}"))
    }

    /// Update access count with instrumentation for consolidation events
    ///
    /// Records MemoryStrengthened events when memories are accessed during retrieval,
    /// capturing activation changes for introspection.
    fn update_access_count_instrumented(&self, memory: &SharedMemory, reason: StrengtheningReason) {
        // Capture activation before update
        let activation_before = memory.importance();

        // Perform the actual access update
        memory.update_access();

        // Capture activation after update
        let activation_after = memory.importance();

        // Only record event if activation actually changed
        if (activation_after - activation_before).abs() > f32::EPSILON {
            let content_preview = if memory.experience.content.chars().count() > 50 {
                let truncated: String = memory.experience.content.chars().take(50).collect();
                format!("{}...", truncated)
            } else {
                memory.experience.content.clone()
            };

            let event = ConsolidationEvent::MemoryStrengthened {
                memory_id: memory.id.0.to_string(),
                content_preview,
                activation_before,
                activation_after,
                reason,
                timestamp: chrono::Utc::now(),
            };

            self.consolidation_events.write().push(event);
        }
    }

    /// Clean up graph episodes for a batch of deleted memory IDs (best-effort)
    fn cleanup_graph_for_ids(&self, ids: &[MemoryId]) {
        if ids.is_empty() {
            return;
        }
        if let Some(graph) = &self.graph_memory {
            let graph_guard = graph.read();
            for id in ids {
                if let Err(e) = graph_guard.delete_episode(&id.0) {
                    tracing::debug!("Graph cleanup failed for {}: {}", &id.0.to_string()[..8], e);
                }
            }
        }
    }

    /// Clean up interference records for a batch of deleted memory IDs (best-effort)
    fn cleanup_interference_for_ids(&self, ids: &[MemoryId]) {
        if ids.is_empty() {
            return;
        }
        let mut detector = self.interference_detector.write();
        for id in ids {
            let id_str = id.0.to_string();
            detector.clear_memory(&id_str);
            if let Err(e) = self.long_term_memory.delete_interference_records(&id_str) {
                tracing::debug!(
                    "Interference cleanup failed for {}: {e}",
                    &id_str[..8.min(id_str.len())]
                );
            }
        }
    }

    /// Forget memories matching a pattern
    ///
    /// Uses validated regex compilation with ReDoS protection
    fn forget_by_pattern(&self, pattern: &str) -> Result<usize> {
        // Use validated pattern compilation with ReDoS protection
        let regex = crate::validation::validate_and_compile_pattern(pattern)
            .map_err(|e| anyhow::anyhow!("Invalid pattern: {e}"))?;
        let mut count = 0;
        let mut working_removed = 0;
        let mut session_removed = 0;
        let mut long_term_removed = 0;

        // Collect IDs from working memory that match
        let working_ids: Vec<MemoryId> = {
            let working = self.working_memory.read();
            working
                .all_memories()
                .iter()
                .filter(|m| regex.is_match(&m.experience.content))
                .map(|m| m.id.clone())
                .collect()
        };
        // Remove from working memory and vector/BM25 index
        {
            let mut working = self.working_memory.write();
            for id in &working_ids {
                if working.remove(id).is_ok() {
                    self.retriever.remove_memory(id);
                    let _ = self.hybrid_search.remove_memory(id);
                    working_removed += 1;
                    count += 1;
                }
            }
        }

        // Collect IDs from session memory that match
        let session_ids: Vec<MemoryId> = {
            let session = self.session_memory.read();
            session
                .all_memories()
                .iter()
                .filter(|m| regex.is_match(&m.experience.content))
                .map(|m| m.id.clone())
                .collect()
        };
        // Remove from session memory and vector/BM25 index
        {
            let mut session = self.session_memory.write();
            for id in &session_ids {
                if session.remove(id).is_ok() {
                    self.retriever.remove_memory(id);
                    let _ = self.hybrid_search.remove_memory(id);
                    session_removed += 1;
                    count += 1;
                }
            }
        }

        // Remove from long-term memory
        let all_lt = self.long_term_memory.get_all()?;
        let mut lt_ids = Vec::new();
        for memory in all_lt {
            if regex.is_match(&memory.experience.content) {
                lt_ids.push(memory.id.clone());
                self.retriever.remove_memory(&memory.id);
                let _ = self.hybrid_search.remove_memory(&memory.id);
                self.long_term_memory.delete(&memory.id)?;
                long_term_removed += 1;
                count += 1;
            }
        }

        // Clean up graph episodes and interference records for all deleted memories
        let all_ids: Vec<MemoryId> = working_ids
            .into_iter()
            .chain(session_ids)
            .chain(lt_ids)
            .collect();
        self.cleanup_graph_for_ids(&all_ids);
        self.cleanup_interference_for_ids(&all_ids);

        // Update stats
        {
            let mut stats = self.stats.write();
            stats.total_memories = stats.total_memories.saturating_sub(count);
            stats.working_memory_count = stats.working_memory_count.saturating_sub(working_removed);
            stats.session_memory_count = stats.session_memory_count.saturating_sub(session_removed);
            stats.long_term_memory_count = stats
                .long_term_memory_count
                .saturating_sub(long_term_removed);
            stats.vector_index_count = stats.vector_index_count.saturating_sub(count);
        }

        Ok(count)
    }

    /// Forget memories matching ANY of the specified tags
    fn forget_by_tags(&self, tags: &[String]) -> Result<usize> {
        let mut count = 0;
        let mut working_removed = 0;
        let mut session_removed = 0;
        let mut long_term_removed = 0;
        let mut all_deleted_ids = Vec::new();

        // Remove from working memory
        {
            let mut working = self.working_memory.write();
            let ids_to_remove: Vec<MemoryId> = working
                .all_memories()
                .iter()
                .filter(|m| m.experience.tags.iter().any(|t| tags.contains(t)))
                .map(|m| m.id.clone())
                .collect();
            for id in &ids_to_remove {
                if working.remove(id).is_ok() {
                    self.retriever.remove_memory(id);
                    let _ = self.hybrid_search.remove_memory(id);
                    working_removed += 1;
                    count += 1;
                }
            }
            all_deleted_ids.extend(ids_to_remove);
        }

        // Remove from session memory
        {
            let mut session = self.session_memory.write();
            let ids_to_remove: Vec<MemoryId> = session
                .all_memories()
                .iter()
                .filter(|m| m.experience.tags.iter().any(|t| tags.contains(t)))
                .map(|m| m.id.clone())
                .collect();
            for id in &ids_to_remove {
                if session.remove(id).is_ok() {
                    self.retriever.remove_memory(id);
                    let _ = self.hybrid_search.remove_memory(id);
                    session_removed += 1;
                    count += 1;
                }
            }
            all_deleted_ids.extend(ids_to_remove);
        }

        // Remove from long-term memory (hard delete for tag-based)
        let all_lt = self.long_term_memory.get_all()?;
        for memory in all_lt {
            if memory.experience.tags.iter().any(|t| tags.contains(t)) {
                all_deleted_ids.push(memory.id.clone());
                self.retriever.remove_memory(&memory.id);
                let _ = self.hybrid_search.remove_memory(&memory.id);
                self.long_term_memory.delete(&memory.id)?;
                long_term_removed += 1;
                count += 1;
            }
        }

        // Clean up graph episodes and interference records for all deleted memories
        self.cleanup_graph_for_ids(&all_deleted_ids);
        self.cleanup_interference_for_ids(&all_deleted_ids);

        // Update stats
        {
            let mut stats = self.stats.write();
            stats.total_memories = stats.total_memories.saturating_sub(count);
            stats.working_memory_count = stats.working_memory_count.saturating_sub(working_removed);
            stats.session_memory_count = stats.session_memory_count.saturating_sub(session_removed);
            stats.long_term_memory_count = stats
                .long_term_memory_count
                .saturating_sub(long_term_removed);
            stats.vector_index_count = stats.vector_index_count.saturating_sub(count);
        }

        Ok(count)
    }

    /// Forget memories within a date range (inclusive)
    fn forget_by_date_range(
        &self,
        start: chrono::DateTime<chrono::Utc>,
        end: chrono::DateTime<chrono::Utc>,
    ) -> Result<usize> {
        let mut count = 0;
        let mut working_removed = 0;
        let mut session_removed = 0;
        let mut long_term_removed = 0;

        // Collect IDs from working memory that match date range
        let working_ids: Vec<MemoryId> = {
            let working = self.working_memory.read();
            working
                .all_memories()
                .iter()
                .filter(|m| m.created_at >= start && m.created_at <= end)
                .map(|m| m.id.clone())
                .collect()
        };
        // Remove from working memory and vector/BM25 index
        {
            let mut working = self.working_memory.write();
            for id in &working_ids {
                if working.remove(id).is_ok() {
                    self.retriever.remove_memory(id);
                    let _ = self.hybrid_search.remove_memory(id);
                    working_removed += 1;
                    count += 1;
                }
            }
        }

        // Collect IDs from session memory that match date range
        let session_ids: Vec<MemoryId> = {
            let session = self.session_memory.read();
            session
                .all_memories()
                .iter()
                .filter(|m| m.created_at >= start && m.created_at <= end)
                .map(|m| m.id.clone())
                .collect()
        };
        // Remove from session memory and vector/BM25 index
        {
            let mut session = self.session_memory.write();
            for id in &session_ids {
                if session.remove(id).is_ok() {
                    self.retriever.remove_memory(id);
                    let _ = self.hybrid_search.remove_memory(id);
                    session_removed += 1;
                    count += 1;
                }
            }
        }

        // Remove from long-term memory using storage search
        let memories = self
            .long_term_memory
            .search(storage::SearchCriteria::ByDate { start, end })?;
        let mut lt_ids = Vec::new();
        for memory in memories {
            lt_ids.push(memory.id.clone());
            self.retriever.remove_memory(&memory.id);
            let _ = self.hybrid_search.remove_memory(&memory.id);
            self.long_term_memory.delete(&memory.id)?;
            long_term_removed += 1;
            count += 1;
        }

        // Clean up graph episodes and interference records for all deleted memories
        let all_ids: Vec<MemoryId> = working_ids
            .into_iter()
            .chain(session_ids)
            .chain(lt_ids)
            .collect();
        self.cleanup_graph_for_ids(&all_ids);
        self.cleanup_interference_for_ids(&all_ids);

        // Update stats
        {
            let mut stats = self.stats.write();
            stats.total_memories = stats.total_memories.saturating_sub(count);
            stats.working_memory_count = stats.working_memory_count.saturating_sub(working_removed);
            stats.session_memory_count = stats.session_memory_count.saturating_sub(session_removed);
            stats.long_term_memory_count = stats
                .long_term_memory_count
                .saturating_sub(long_term_removed);
            stats.vector_index_count = stats.vector_index_count.saturating_sub(count);
        }

        Ok(count)
    }

    /// Forget memories of a specific type
    fn forget_by_type(&self, exp_type: ExperienceType) -> Result<usize> {
        let mut count = 0;
        let mut working_removed = 0;
        let mut session_removed = 0;
        let mut long_term_removed = 0;

        // Collect IDs from working memory that match type
        let working_ids: Vec<MemoryId> = {
            let working = self.working_memory.read();
            working
                .all_memories()
                .iter()
                .filter(|m| m.experience.experience_type == exp_type)
                .map(|m| m.id.clone())
                .collect()
        };
        // Remove from working memory and vector/BM25 index
        {
            let mut working = self.working_memory.write();
            for id in &working_ids {
                if working.remove(id).is_ok() {
                    self.retriever.remove_memory(id);
                    let _ = self.hybrid_search.remove_memory(id);
                    working_removed += 1;
                    count += 1;
                }
            }
        }

        // Collect IDs from session memory that match type
        let session_ids: Vec<MemoryId> = {
            let session = self.session_memory.read();
            session
                .all_memories()
                .iter()
                .filter(|m| m.experience.experience_type == exp_type)
                .map(|m| m.id.clone())
                .collect()
        };
        // Remove from session memory and vector/BM25 index
        {
            let mut session = self.session_memory.write();
            for id in &session_ids {
                if session.remove(id).is_ok() {
                    self.retriever.remove_memory(id);
                    let _ = self.hybrid_search.remove_memory(id);
                    session_removed += 1;
                    count += 1;
                }
            }
        }

        // Remove from long-term memory using storage search
        let memories = self
            .long_term_memory
            .search(storage::SearchCriteria::ByType(exp_type))?;
        let mut lt_ids = Vec::new();
        for memory in memories {
            lt_ids.push(memory.id.clone());
            self.retriever.remove_memory(&memory.id);
            let _ = self.hybrid_search.remove_memory(&memory.id);
            self.long_term_memory.delete(&memory.id)?;
            long_term_removed += 1;
            count += 1;
        }

        // Clean up graph episodes and interference records for all deleted memories
        let all_ids: Vec<MemoryId> = working_ids
            .into_iter()
            .chain(session_ids)
            .chain(lt_ids)
            .collect();
        self.cleanup_graph_for_ids(&all_ids);
        self.cleanup_interference_for_ids(&all_ids);

        // Update stats
        {
            let mut stats = self.stats.write();
            stats.total_memories = stats.total_memories.saturating_sub(count);
            stats.working_memory_count = stats.working_memory_count.saturating_sub(working_removed);
            stats.session_memory_count = stats.session_memory_count.saturating_sub(session_removed);
            stats.long_term_memory_count = stats
                .long_term_memory_count
                .saturating_sub(long_term_removed);
            stats.vector_index_count = stats.vector_index_count.saturating_sub(count);
        }

        Ok(count)
    }

    /// Forget ALL memories for a user (GDPR compliance - right to erasure)
    ///
    /// WARNING: This is a destructive operation. All memories across all tiers
    /// will be permanently deleted. This cannot be undone.
    fn forget_all(&self) -> Result<usize> {
        // Deletion order: graph → long-term → session → working → stats
        // This is fail-safe: if we crash mid-way, the most durable data
        // (graph/long-term) is already deleted. Working/session memory is
        // ephemeral and will be empty on restart anyway.

        let mut count = 0;

        // Step 1: Clear knowledge graph first (GDPR - complete erasure)
        // Graph references memories, so clean references before deleting memories
        if let Some(graph) = &self.graph_memory {
            match graph.read().clear_all() {
                Ok((entities, relationships, episodes)) => {
                    tracing::info!(
                        entities,
                        relationships,
                        episodes,
                        "Graph cleared during forget_all"
                    );
                }
                Err(e) => {
                    tracing::warn!(error = %e, "Failed to clear knowledge graph during forget_all");
                    // Continue — graph cleanup is best-effort for GDPR
                }
            }
        }

        // Step 2: Clear long-term memory (persistent, most important to delete)
        let all_lt = self.long_term_memory.get_all()?;
        let long_term_count = all_lt.len();
        for memory in all_lt {
            self.retriever.remove_memory(&memory.id);
            let _ = self.hybrid_search.remove_memory(&memory.id);
            self.long_term_memory.delete(&memory.id)?;
        }
        count += long_term_count;

        // Step 3: Clear session memory (ephemeral — lost on restart anyway)
        let session_ids: Vec<MemoryId> = {
            let session = self.session_memory.read();
            session
                .all_memories()
                .iter()
                .map(|m| m.id.clone())
                .collect()
        };
        let session_count = session_ids.len();
        for id in &session_ids {
            self.retriever.remove_memory(id);
            let _ = self.hybrid_search.remove_memory(id);
        }
        {
            let mut session = self.session_memory.write();
            session.clear();
        }
        count += session_count;

        // Step 4: Clear working memory (ephemeral — lost on restart anyway)
        let working_ids: Vec<MemoryId> = {
            let working = self.working_memory.read();
            working
                .all_memories()
                .iter()
                .map(|m| m.id.clone())
                .collect()
        };
        let working_count = working_ids.len();
        for id in &working_ids {
            self.retriever.remove_memory(id);
            let _ = self.hybrid_search.remove_memory(id);
        }
        {
            let mut working = self.working_memory.write();
            working.clear();
        }
        count += working_count;

        // Step 5: Commit BM25 deletions
        if let Err(e) = self.hybrid_search.commit_and_reload() {
            tracing::warn!(error = %e, "BM25 commit failed during forget_all");
        }

        // Step 6: Clear semantic facts (GDPR — knowledge derived from memories)
        {
            let db = self.long_term_memory.db();
            let mut batch = rocksdb::WriteBatch::default();
            let mut facts_deleted = 0usize;
            for prefix in &[
                "facts:",
                "facts_by_entity:",
                "facts_by_type:",
                "facts_embedding:",
            ] {
                let iter = db.prefix_iterator(prefix.as_bytes());
                for item in iter {
                    if let Ok((key, _)) = item {
                        if !key.starts_with(prefix.as_bytes()) {
                            break;
                        }
                        batch.delete(&key);
                        if *prefix == "facts:" {
                            facts_deleted += 1;
                        }
                    }
                }
            }
            // Clear temporal facts
            let iter = db.prefix_iterator(b"temporal_facts:");
            for item in iter {
                if let Ok((key, _)) = item {
                    if !key.starts_with(b"temporal_facts:") {
                        break;
                    }
                    batch.delete(&key);
                }
            }
            if facts_deleted > 0 || !batch.is_empty() {
                if let Err(e) = db.write(batch) {
                    tracing::warn!(error = %e, "Failed to clear facts during forget_all");
                } else {
                    tracing::info!(facts_deleted, "Semantic facts cleared during forget_all");
                }
            }
        }

        // Step 7: Clear interference history (in-memory + persisted)
        {
            let mut detector = self.interference_detector.write();
            *detector = replay::InterferenceDetector::new();
        }
        if let Err(e) = self.long_term_memory.clear_all_interference_records() {
            tracing::warn!(error = %e, "Failed to clear interference records during forget_all");
        }

        // Step 8: Reset stats last (reflects final state)
        {
            let mut stats = self.stats.write();
            stats.total_memories = 0;
            stats.working_memory_count = 0;
            stats.session_memory_count = 0;
            stats.long_term_memory_count = 0;
            stats.vector_index_count = 0;
        }

        Ok(count)
    }

    /// Show memory visualization (ASCII art of memory graph)
    pub fn show_visualization(&self) {
        self.logger.read().show_visualization();
    }

    /// Export memory graph as DOT file for Graphviz
    pub fn export_graph(&self, path: &std::path::Path) -> Result<()> {
        self.logger.read().export_dot(path)
    }

    /// Get visualization statistics
    pub fn get_visualization_stats(&self) -> GraphStats {
        self.logger.read().get_stats()
    }

    /// Flush long-term storage to ensure data persistence (critical for graceful shutdown)
    pub fn flush_storage(&self) -> Result<()> {
        // Flush RocksDB storage
        self.long_term_memory.flush()?;

        // Persist vector index and ID mapping for restart recovery
        self.retriever.save()?;

        Ok(())
    }

    /// Get the underlying RocksDB database handle for backup operations
    ///
    /// # Warning
    /// This provides direct access to the database. Use with caution.
    /// Primarily intended for backup/restore operations.
    pub fn get_db(&self) -> std::sync::Arc<rocksdb::DB> {
        self.long_term_memory.db()
    }

    /// Advanced search using storage criteria
    pub fn advanced_search(&self, criteria: storage::SearchCriteria) -> Result<Vec<Memory>> {
        self.long_term_memory.search(criteria)
    }

    /// Get memory by ID from long-term storage
    pub fn get_memory(&self, id: &MemoryId) -> Result<Memory> {
        self.long_term_memory.get(id)
    }

    /// Update a memory in storage with full re-indexing
    ///
    /// This properly updates the memory by:
    /// 1. Removing stale secondary indices and re-storing in RocksDB
    /// 2. Re-indexing in vector DB (semantic search) if embeddings changed
    /// 3. Re-indexing in BM25 (keyword/hybrid search)
    /// 4. Updating working/session memory caches if the memory is cached
    pub fn update_memory(&self, memory: &Memory) -> Result<()> {
        let memory_id = memory.id.clone();

        // Update in storage (removes old indices, re-stores with fresh indices)
        self.long_term_memory.update(memory)?;

        // Re-index in vector DB with updated embeddings
        if let Err(e) = self.retriever.reindex_memory(memory) {
            tracing::warn!(
                "Failed to reindex memory {} in vector DB: {}",
                memory_id.0,
                e
            );
        }

        // Re-index in BM25 with updated content
        if let Err(e) = self.hybrid_search.index_memory(
            &memory_id,
            &memory.experience.content,
            &memory.experience.tags,
            &memory.experience.entities,
        ) {
            tracing::warn!("Failed to reindex memory {} in BM25: {}", memory_id.0, e);
        }
        if let Err(e) = self.hybrid_search.commit_and_reload() {
            tracing::warn!("Failed to commit/reload BM25 index: {}", e);
        }

        // Update in working/session memory caches if present
        {
            let mut working = self.working_memory.write();
            if working.contains(&memory_id) {
                let _ = working.remove(&memory_id);
                let _ = working.add_shared(std::sync::Arc::new(memory.clone()));
            }
        }
        {
            let mut session = self.session_memory.write();
            if session.contains(&memory_id) {
                let _ = session.remove(&memory_id);
                let _ = session.add_shared(std::sync::Arc::new(memory.clone()));
            }
        }

        Ok(())
    }

    /// Set or update the parent of a memory for hierarchical organization
    ///
    /// This enables memory trees where memories can have parent-child relationships.
    /// Example: "71-research" -> "algebraic" -> "21×27≡-1"
    ///
    /// Pass `None` as parent_id to remove the parent (make it a root memory).
    pub fn set_memory_parent(
        &self,
        memory_id: &MemoryId,
        parent_id: Option<MemoryId>,
    ) -> Result<()> {
        // Update the persistent copy in long-term storage
        let mut memory = self.long_term_memory.get(memory_id)?;
        memory.set_parent(parent_id.clone());
        self.long_term_memory.update(&memory)?;

        // Also update the in-memory tier copy (working or session) so reads
        // reflect the parent_id immediately without waiting for tier promotion
        let updated = Arc::new(memory);
        {
            let mut wm = self.working_memory.write();
            if wm.contains(memory_id) {
                let _ = wm.remove(memory_id);
                let _ = wm.add_shared(Arc::clone(&updated));
            }
        }
        {
            let mut sm = self.session_memory.write();
            if sm.contains(memory_id) {
                let _ = sm.remove(memory_id);
                let _ = sm.add_shared(Arc::clone(&updated));
            }
        }

        Ok(())
    }

    /// Get children of a memory
    pub fn get_memory_children(&self, parent_id: &MemoryId) -> Result<Vec<Memory>> {
        self.long_term_memory.get_children(parent_id)
    }

    /// Get ancestors (parent chain) of a memory
    pub fn get_memory_ancestors(&self, memory_id: &MemoryId) -> Result<Vec<Memory>> {
        self.long_term_memory.get_ancestors(memory_id)
    }

    /// Get full hierarchy context (ancestors, memory, children)
    pub fn get_memory_hierarchy(
        &self,
        memory_id: &MemoryId,
    ) -> Result<(Vec<Memory>, Memory, Vec<Memory>)> {
        self.long_term_memory.get_hierarchy_context(memory_id)
    }

    /// Decompress a memory
    pub fn decompress_memory(&self, memory: &Memory) -> Result<Memory> {
        self.compressor.decompress(memory)
    }

    /// Get storage statistics
    pub fn get_storage_stats(&self) -> Result<storage::StorageStats> {
        self.long_term_memory.get_stats()
    }

    /// Get uncompressed old memories
    pub fn get_uncompressed_older_than(
        &self,
        cutoff: chrono::DateTime<chrono::Utc>,
    ) -> Result<Vec<Memory>> {
        self.long_term_memory.get_uncompressed_older_than(cutoff)
    }

    /// Rebuild vector index from all existing long-term memories (startup initialization)
    pub fn rebuild_vector_index(&self) -> Result<()> {
        self.retriever.rebuild_index()
    }

    /// Repair vector index by finding and re-indexing orphaned memories
    ///
    /// Orphaned memories are those stored in RocksDB but missing from the vector index.
    /// This can happen if embedding generation fails during record().
    ///
    /// Returns: (total_storage, indexed, repaired, failed)
    pub fn repair_vector_index(&self) -> Result<(usize, usize, usize, usize)> {
        let all_memories = self.long_term_memory.get_all()?;
        let total_storage = all_memories.len();
        let indexed_before = self.retriever.len();

        let mut repaired = 0;
        let mut failed = 0;

        // Get set of indexed memory IDs
        let indexed_ids = self.retriever.get_indexed_memory_ids();

        for memory in all_memories {
            // Check if memory is already indexed
            if indexed_ids.contains(&memory.id) {
                continue;
            }

            // Memory is orphaned - try to index it
            tracing::info!(
                memory_id = %memory.id.0,
                content_preview = %memory.experience.content.chars().take(50).collect::<String>(),
                "Repairing orphaned memory"
            );

            match self.retriever.index_memory(&memory) {
                Ok(_) => {
                    repaired += 1;
                    tracing::info!(memory_id = %memory.id.0, "Successfully repaired orphaned memory");
                }
                Err(e) => {
                    failed += 1;
                    tracing::error!(
                        memory_id = %memory.id.0,
                        error = %e,
                        "Failed to repair orphaned memory - embedding generation failed"
                    );
                }
            }
        }

        let indexed_after = self.retriever.len();

        // Update stats
        {
            let mut stats = self.stats.write();
            stats.vector_index_count = indexed_after;
        }

        tracing::info!(
            total_storage = total_storage,
            indexed_before = indexed_before,
            indexed_after = indexed_after,
            repaired = repaired,
            failed = failed,
            "Vector index repair completed"
        );

        Ok((total_storage, indexed_after, repaired, failed))
    }

    /// Verify index integrity and return diagnostic information
    ///
    /// Returns a struct with:
    /// - total_storage: memories in RocksDB
    /// - total_indexed: memories in vector index
    /// - orphaned_count: memories missing from index
    /// - orphaned_ids: list of orphaned memory IDs (first 100)
    pub fn verify_index_integrity(&self) -> Result<IndexIntegrityReport> {
        let all_memories = self.long_term_memory.get_all()?;
        let total_storage = all_memories.len();
        let indexed_ids = self.retriever.get_indexed_memory_ids();
        let total_indexed = indexed_ids.len();

        let mut orphaned_ids = Vec::new();
        for memory in &all_memories {
            if !indexed_ids.contains(&memory.id) {
                if orphaned_ids.len() < 100 {
                    orphaned_ids.push(memory.id.clone());
                }
            }
        }

        let orphaned_count = total_storage.saturating_sub(total_indexed);

        let is_healthy = orphaned_count == 0;
        Ok(IndexIntegrityReport {
            total_storage,
            total_indexed,
            orphaned_count,
            orphaned_ids,
            is_healthy,
            healthy: is_healthy,
        })
    }

    /// Cleanup corrupted memories that fail to deserialize
    /// Returns the number of entries deleted
    pub fn cleanup_corrupted(&self) -> Result<usize> {
        self.long_term_memory.cleanup_corrupted()
    }

    /// Migrate legacy memories to current format for improved performance
    /// Returns (migrated_count, already_current_count, failed_count)
    pub fn migrate_legacy(&self) -> Result<(usize, usize, usize)> {
        self.long_term_memory.migrate_legacy()
    }

    /// Rebuild vector index from scratch using only valid memories in storage
    /// This removes orphaned index entries and rebuilds with proper ID mappings
    /// Returns (total_memories, total_indexed)
    pub fn rebuild_index(&self) -> Result<(usize, usize)> {
        tracing::info!("Starting full index rebuild from storage");
        self.retriever.rebuild_index()?;
        let indexed = self.retriever.len();
        let storage_count = self.long_term_memory.get_stats()?.total_count;

        // Update stats
        {
            let mut stats = self.stats.write();
            stats.vector_index_count = indexed;
        }

        tracing::info!(
            storage_count = storage_count,
            indexed = indexed,
            "Index rebuild complete"
        );

        Ok((storage_count, indexed))
    }

    /// Save vector index to disk (shutdown persistence)
    /// Uses Vamana persistence format for instant startup on restart
    pub fn save_vector_index(&self, _path: &Path) -> Result<()> {
        self.retriever.save()
    }

    /// Merge BM25 index segments to remove ghost state from overwrites.
    /// Should be called during heavy maintenance cycles only (expensive operation).
    /// Returns the number of segments merged (0 if already optimal).
    pub fn optimize_bm25(&self) -> Result<usize> {
        self.hybrid_search.optimize_bm25()
    }

    /// Drain the write retry buffer, re-attempting any failed stores.
    /// Returns the number of successfully retried writes.
    pub fn drain_write_retries(&self) -> usize {
        self.long_term_memory.drain_retry_buffer()
    }

    /// Number of writes pending retry
    pub fn pending_write_retries(&self) -> usize {
        self.long_term_memory.pending_retry_count()
    }

    /// Total write failures since server start
    pub fn total_write_failures(&self) -> u64 {
        self.long_term_memory.total_write_failures()
    }

    /// Get BM25 segment count for health metrics
    pub fn bm25_segment_count(&self) -> usize {
        self.hybrid_search.bm25_segment_count()
    }

    /// Get vector index health information
    ///
    /// Returns metrics about the Vamana index including total vectors,
    /// incremental inserts since last build, and whether rebuild is recommended.
    pub fn index_health(&self) -> retrieval::IndexHealth {
        self.retriever.index_health()
    }

    /// Auto-rebuild vector index if degradation threshold is exceeded
    ///
    /// Returns `Ok(true)` if rebuild was performed, `Ok(false)` if not needed.
    /// Thread-safe: concurrent calls are no-ops while rebuild is in progress.
    pub fn auto_rebuild_index_if_needed(&self) -> Result<bool> {
        self.retriever.auto_rebuild_index_if_needed()
    }

    /// Auto-repair index integrity and compact if needed
    ///
    /// Called during maintenance to ensure storage↔index consistency:
    /// 1. Checks index health (fast O(1) operation)
    /// 2. If needs compaction (>30% deleted), triggers auto-rebuild
    /// 3. If orphaned memories detected, repairs them
    ///
    /// This provides eventual consistency between storage and index.
    fn auto_repair_and_compact(&self) {
        // Check index health first (fast operation)
        let health = self.index_health();

        // Auto-compact if deletion ratio exceeds threshold
        if health.needs_compaction {
            tracing::info!(
                "Index compaction triggered: {:.1}% deleted ({} of {} vectors)",
                health.deletion_ratio * 100.0,
                health.deleted_count,
                health.total_vectors
            );
            if let Err(e) = self.auto_rebuild_index_if_needed() {
                tracing::warn!("Index compaction failed: {}", e);
            }
        }

        // Check for orphaned memories (stored but not indexed)
        // Only do full scan if we suspect issues (cheap heuristic: counts differ)
        let storage_count = self
            .long_term_memory
            .get_stats()
            .map(|s| s.total_count)
            .unwrap_or(0);
        let index_count = health.total_vectors.saturating_sub(health.deleted_count);

        if storage_count > index_count {
            // Potential orphans detected - run repair
            let orphan_estimate = storage_count - index_count;
            if orphan_estimate > 0 {
                tracing::info!(
                    "Potential orphaned memories detected: ~{} (storage={}, indexed={})",
                    orphan_estimate,
                    storage_count,
                    index_count
                );
                match self.repair_vector_index() {
                    Ok((_, _, repaired, failed)) => {
                        if repaired > 0 || failed > 0 {
                            tracing::info!(
                                "Index repair complete: {} repaired, {} failed",
                                repaired,
                                failed
                            );
                        }
                    }
                    Err(e) => {
                        tracing::warn!("Index repair failed: {}", e);
                    }
                }
            }
        }
    }

    // =========================================================================
    // OUTCOME FEEDBACK SYSTEM - Hebbian "Fire Together, Wire Together"
    // =========================================================================

    /// Retrieve memories with tracking for later feedback
    ///
    /// Use this when you want to provide feedback on retrieval quality.
    /// Returns a TrackedRetrieval that can be used with `reinforce_recall`.
    ///
    /// # Example
    /// ```ignore
    /// let tracked = memory_system.recall_tracked(&query)?;
    /// // Use memories...
    /// // Later, after task completion:
    /// memory_system.reinforce_recall(&tracked.memory_ids(), RetrievalOutcome::Helpful)?;
    /// ```
    pub fn recall_tracked(&self, query: &Query) -> Result<TrackedRetrieval> {
        let result = self.retriever.search_tracked(query, query.max_results)?;
        if let Ok(count) = self.long_term_memory.increment_retrieval_count() {
            self.stats.write().total_retrievals = count;
        }
        Ok(result)
    }

    /// Reinforce memories based on task outcome (core feedback loop)
    ///
    /// This is THE key method that closes the Hebbian loop:
    /// - If outcome is Helpful: strengthen associations, boost importance
    /// - If outcome is Misleading: weaken associations, reduce importance
    /// - If outcome is Neutral: just record access (mild reinforcement)
    ///
    /// CACHE COHERENCY: This method updates BOTH the in-memory caches AND
    /// persistent storage to ensure importance changes are visible immediately
    /// through cached references (via Arc interior mutability) AND survive restarts.
    ///
    /// # Arguments
    /// * `memory_ids` - IDs of memories that were used in the task
    /// * `outcome` - Whether the memories were helpful, misleading, or neutral
    ///
    /// # Returns
    /// Statistics about what was reinforced
    pub fn reinforce_recall(
        &self,
        memory_ids: &[MemoryId],
        outcome: RetrievalOutcome,
    ) -> Result<ReinforcementStats> {
        if memory_ids.is_empty() {
            return Ok(ReinforcementStats::default());
        }

        let mut stats = ReinforcementStats {
            memories_processed: memory_ids.len(),
            outcome: outcome.clone(),
            ..Default::default()
        };

        // Hebbian coactivation: strengthen associations between co-retrieved memories
        // Uses GraphMemory if available, otherwise counts pair associations directly
        if !matches!(outcome, RetrievalOutcome::Misleading) && memory_ids.len() >= 2 {
            if let Some(graph) = &self.graph_memory {
                let memory_uuids: Vec<uuid::Uuid> = memory_ids.iter().map(|id| id.0).collect();
                if memory_uuids.len() >= 2 {
                    match graph.read().record_memory_coactivation(&memory_uuids) {
                        Ok(count) => {
                            stats.associations_strengthened = count;
                        }
                        Err(e) => {
                            tracing::warn!(error = %e, "Failed to record memory coactivation");
                            // Fallback: count pairs
                            let n = memory_ids.len();
                            stats.associations_strengthened = n * (n - 1) / 2;
                        }
                    }
                }
            } else {
                // No graph memory available — count pairs directly
                let n = memory_ids.len();
                stats.associations_strengthened = n * (n - 1) / 2;
            }
        }

        // Hebbian feedback on entity-level graph edges:
        // Helpful → strengthen entity edges for involved memories
        // Misleading → weaken entity edges (anti-Hebbian)
        if let Some(graph) = &self.graph_memory {
            let memory_uuids: Vec<uuid::Uuid> = memory_ids.iter().map(|id| id.0).collect();
            const MAX_FEEDBACK_EDGES: usize = 200;
            match graph
                .read()
                .collect_entity_edges_for_memories(&memory_uuids, MAX_FEEDBACK_EDGES)
            {
                Ok(edge_uuids) if !edge_uuids.is_empty() => {
                    let result = match outcome {
                        RetrievalOutcome::Helpful => {
                            graph.read().batch_strengthen_synapses(&edge_uuids)
                        }
                        RetrievalOutcome::Misleading => graph
                            .read()
                            .batch_weaken_synapses(&edge_uuids, HEBBIAN_DECAY_MISLEADING),
                        RetrievalOutcome::Neutral => Ok(0),
                    };
                    match result {
                        Ok(count) => {
                            stats.entity_edges_reinforced = count;
                        }
                        Err(e) => {
                            tracing::warn!(
                                error = %e,
                                "Failed to reinforce entity-level graph edges"
                            );
                        }
                    }
                }
                Ok(_) => {} // No entity edges found
                Err(e) => {
                    tracing::warn!(
                        error = %e,
                        "Failed to collect entity edges for feedback reinforcement"
                    );
                }
            }
        }

        // CACHE COHERENT IMPORTANCE UPDATES:
        // 1. First try to find memory in caches (working, session)
        // 2. If found in cache, modify through the cached Arc (interior mutability)
        //    This updates ALL holders of this Arc reference
        // 3. Then persist to storage for durability
        // 4. If not in cache, get from storage, modify, and persist
        let mut persist_failures: Vec<(MemoryId, String)> = Vec::new();
        let mut total_prediction_error = 0.0f32;
        let mut prediction_error_count = 0u32;

        for id in memory_ids {
            // Prediction error weighting (VTA/Dopamine system — Schultz 1997):
            // Look up how important we predicted this memory to be when surfaced.
            // Compute prediction error → scale learning signal by surprise level.
            let predicted = self.prediction_cache.get(id).unwrap_or(0.5);
            let prediction_error = match &outcome {
                RetrievalOutcome::Helpful => 1.0 - predicted, // High importance + Helpful = expected = low error
                RetrievalOutcome::Misleading => predicted, // High importance + Misleading = surprising = high error
                RetrievalOutcome::Neutral => 0.5,          // Baseline
            };
            let error_multiplier = crate::constants::PREDICTION_ERROR_MIN_MULTIPLIER
                + prediction_error
                    * (crate::constants::PREDICTION_ERROR_MAX_MULTIPLIER
                        - crate::constants::PREDICTION_ERROR_MIN_MULTIPLIER);
            total_prediction_error += prediction_error;
            prediction_error_count += 1;

            // Try working memory cache first
            let cached_memory = {
                let working = self.working_memory.read();
                working.get(id)
            };

            // Try session memory cache if not in working
            let cached_memory = cached_memory.or_else(|| {
                let session = self.session_memory.read();
                session.get(id)
            });

            if let Some(memory) = cached_memory {
                // CACHE HIT: Modify through cached Arc (updates all references)
                memory.record_access();
                match &outcome {
                    RetrievalOutcome::Helpful => {
                        memory.boost_importance(HEBBIAN_BOOST_HELPFUL * error_multiplier);
                        stats.importance_boosts += 1;
                    }
                    RetrievalOutcome::Misleading => {
                        memory.decay_importance(HEBBIAN_DECAY_MISLEADING * error_multiplier);
                        stats.importance_decays += 1;
                    }
                    RetrievalOutcome::Neutral => {
                        // Just access recorded
                    }
                }
                // PERSIST: Write updated memory to durable storage
                // Track failures instead of silently ignoring
                if let Err(e) = self.long_term_memory.update(&memory) {
                    persist_failures.push((id.clone(), e.to_string()));
                    tracing::warn!(
                        memory_id = %id.0,
                        error = %e,
                        "Failed to persist reinforcement update - Hebbian feedback may be lost on restart"
                    );
                }
            } else {
                // CACHE MISS: Get from storage, modify, and persist
                match self.long_term_memory.get(id) {
                    Ok(memory) => {
                        memory.record_access();
                        match &outcome {
                            RetrievalOutcome::Helpful => {
                                memory.boost_importance(HEBBIAN_BOOST_HELPFUL * error_multiplier);
                                stats.importance_boosts += 1;
                            }
                            RetrievalOutcome::Misleading => {
                                memory
                                    .decay_importance(HEBBIAN_DECAY_MISLEADING * error_multiplier);
                                stats.importance_decays += 1;
                            }
                            RetrievalOutcome::Neutral => {
                                // Just access recorded
                            }
                        }
                        // PERSIST: Write to durable storage
                        if let Err(e) = self.long_term_memory.update(&memory) {
                            persist_failures.push((id.clone(), e.to_string()));
                            tracing::warn!(
                                memory_id = %id.0,
                                error = %e,
                                "Failed to persist reinforcement update - Hebbian feedback may be lost on restart"
                            );
                        }
                    }
                    Err(e) => {
                        tracing::debug!(
                            memory_id = %id.0,
                            error = %e,
                            "Memory not found during reinforcement - may have been deleted"
                        );
                    }
                }
            }
        }

        // Report aggregate persistence failures
        if !persist_failures.is_empty() {
            stats.persist_failures = persist_failures.len();
            tracing::error!(
                failure_count = persist_failures.len(),
                "Hebbian reinforcement had persistence failures - learning feedback partially lost"
            );
        }

        // Record average prediction error multiplier for observability
        if prediction_error_count > 0 {
            let avg_error = total_prediction_error / prediction_error_count as f32;
            stats.prediction_error_multiplier = crate::constants::PREDICTION_ERROR_MIN_MULTIPLIER
                + avg_error
                    * (crate::constants::PREDICTION_ERROR_MAX_MULTIPLIER
                        - crate::constants::PREDICTION_ERROR_MIN_MULTIPLIER);
        }

        Ok(stats)
    }

    /// Reinforce using a tracked recall (convenience wrapper)
    pub fn reinforce_recall_tracked(
        &self,
        tracked: &TrackedRetrieval,
        outcome: RetrievalOutcome,
    ) -> Result<ReinforcementStats> {
        self.retriever.reinforce_tracked(tracked, outcome)
    }

    /// Perform graph maintenance (decay old edges, prune weak ones)
    ///
    /// Call this periodically (e.g., every hour or on user logout)
    /// to let unused associations naturally fade.
    pub fn graph_maintenance(&self) {
        if let Some(graph) = &self.graph_memory {
            if let Err(e) = graph.read().apply_decay() {
                tracing::debug!("Graph decay maintenance failed: {e}");
            }
        }
    }

    /// Connect extracted facts to the knowledge graph.
    ///
    /// For each fact, ensures all related entities exist as EntityNodes and creates
    /// RelationshipEdges between all pairs. Uses L2Episodic tier because facts are
    /// consolidated knowledge (survived 7-day aging + 2+ supporting memories).
    /// `add_entity` upserts (increments mention_count if existing), and
    /// `add_relationship` strengthens via Hebbian learning if the edge already exists.
    fn connect_facts_to_graph(&self, facts: &[SemanticFact]) {
        let graph = match &self.graph_memory {
            Some(g) => g,
            None => return,
        };
        let graph_guard = graph.read();
        let now = chrono::Utc::now();
        let mut entities_added = 0;
        let mut edges_added = 0;

        // Collect all unique entity names across all facts for batch encoding
        let mut all_entity_names: Vec<String> = Vec::new();
        for fact in facts {
            for name in &fact.related_entities {
                if !all_entity_names.contains(name) {
                    all_entity_names.push(name.clone());
                }
            }
        }

        // Batch-encode entity names for concept-level dedup
        let embedding_map: std::collections::HashMap<String, Vec<f32>> =
            if all_entity_names.is_empty() {
                std::collections::HashMap::new()
            } else {
                let name_refs: Vec<&str> = all_entity_names.iter().map(|s| s.as_str()).collect();
                match self.embedder.encode_batch(&name_refs) {
                    Ok(embs) => all_entity_names.into_iter().zip(embs).collect(),
                    Err(e) => {
                        tracing::debug!(
                            error = %e,
                            "Fact entity name embedding failed, concept merge disabled"
                        );
                        std::collections::HashMap::new()
                    }
                }
            };

        for fact in facts {
            // Ensure all related entities exist as graph nodes
            for entity_name in &fact.related_entities {
                let entity = crate::graph_memory::EntityNode {
                    uuid: Uuid::new_v4(),
                    name: entity_name.clone(),
                    labels: vec![crate::graph_memory::EntityLabel::Concept],
                    created_at: now,
                    last_seen_at: now,
                    mention_count: 1,
                    summary: String::new(),
                    attributes: std::collections::HashMap::new(),
                    name_embedding: embedding_map.get(entity_name).cloned(),
                    salience: fact.confidence * 0.5,
                    is_proper_noun: entity_name
                        .chars()
                        .next()
                        .map(|c| c.is_uppercase())
                        .unwrap_or(false),
                    selectivity: None,
                };
                if graph_guard.add_entity(entity).is_ok() {
                    entities_added += 1;
                }
            }

            // Create edges between all pairs of related entities with semantic weighting
            let entities = &fact.related_entities;
            let l2_base_weight = crate::graph_memory::EdgeTier::L2Episodic.initial_weight();
            for i in 0..entities.len() {
                for j in (i + 1)..entities.len() {
                    if let (Ok(Some(e1)), Ok(Some(e2))) = (
                        graph_guard.find_entity_by_name(&entities[i]),
                        graph_guard.find_entity_by_name(&entities[j]),
                    ) {
                        let semantic_weight = match (&e1.name_embedding, &e2.name_embedding) {
                            (Some(emb1), Some(emb2)) => {
                                let sim = crate::similarity::cosine_similarity(emb1, emb2).max(0.0);
                                EDGE_SEMANTIC_WEIGHT_FLOOR
                                    + (1.0 - EDGE_SEMANTIC_WEIGHT_FLOOR) * sim
                            }
                            _ => 1.0,
                        };

                        let edge = crate::graph_memory::RelationshipEdge {
                            uuid: Uuid::new_v4(),
                            from_entity: e1.uuid,
                            to_entity: e2.uuid,
                            relation_type: crate::graph_memory::RelationType::RelatedTo,
                            strength: l2_base_weight * semantic_weight,
                            created_at: now,
                            valid_at: now,
                            invalidated_at: None,
                            source_episode_id: None,
                            context: fact.fact.chars().take(100).collect(),
                            last_activated: now,
                            activation_count: 1,
                            ltp_status: crate::graph_memory::LtpStatus::None,
                            tier: crate::graph_memory::EdgeTier::L2Episodic,
                            activation_timestamps: None,
                            entity_confidence: Some(fact.confidence),
                            forman_curvature: None,
                            endpoint_selectivity: None,
                        };
                        if graph_guard.add_relationship(edge).is_ok() {
                            edges_added += 1;
                        }
                    }
                }
            }
        }

        if entities_added > 0 || edges_added > 0 {
            tracing::debug!(
                entities_added,
                edges_added,
                facts = facts.len(),
                "Connected facts to knowledge graph"
            );
        }
    }

    /// Get memory graph statistics
    pub fn graph_stats(&self) -> MemoryGraphStats {
        if let Some(graph) = &self.graph_memory {
            let g = graph.read();
            if let Ok(stats) = g.get_stats() {
                // Calculate avg_strength and potentiated_count from relationships
                let (avg_strength, potentiated_count) = if let Ok(relationships) =
                    g.get_all_relationships()
                {
                    if relationships.is_empty() {
                        (0.0, 0)
                    } else {
                        let total_strength: f32 = relationships.iter().map(|r| r.strength).sum();
                        let potentiated =
                            relationships.iter().filter(|r| r.is_potentiated()).count();
                        (total_strength / relationships.len() as f32, potentiated)
                    }
                } else {
                    (0.0, 0)
                };

                return MemoryGraphStats {
                    node_count: stats.entity_count,
                    edge_count: stats.relationship_count,
                    avg_strength,
                    potentiated_count,
                };
            }
        }
        // Return empty stats if no graph
        MemoryGraphStats {
            node_count: 0,
            edge_count: 0,
            avg_strength: 0.0,
            potentiated_count: 0,
        }
    }

    // =========================================================================
    // UPSERT: Mutable memories with external linking and audit history
    // =========================================================================

    /// Upsert a memory: create if new, update with history tracking if exists
    ///
    /// When a memory with the same external_id exists:
    /// 1. Old content is pushed to history (audit trail)
    /// 2. Content is updated with new content
    /// 3. Version is incremented
    /// 4. Embeddings are regenerated for new content
    /// 5. Vector index is updated
    ///
    /// # Arguments
    /// * `external_id` - External system identifier (e.g., "linear:SHO-39", "github:pr-123")
    /// * `experience` - The experience data to store
    /// * `change_type` - Type of change (ContentUpdated, StatusChanged, etc.)
    /// * `changed_by` - Optional: who/what triggered the change
    /// * `change_reason` - Optional: description of why this changed
    ///
    /// # Returns
    /// * `(MemoryId, bool)` - Memory ID and whether it was an update (true) or create (false)
    pub fn upsert(
        &self,
        external_id: String,
        mut experience: Experience,
        change_type: ChangeType,
        changed_by: Option<String>,
        change_reason: Option<String>,
    ) -> Result<(MemoryId, bool)> {
        // Try to find existing memory with this external_id
        if let Some(mut existing) = self.long_term_memory.find_by_external_id(&external_id)? {
            // === UPDATE PATH ===
            let memory_id = existing.id.clone();

            // Push old content to history and update
            existing.update_content(
                experience.content.clone(),
                change_type,
                changed_by,
                change_reason,
            );

            // Update entities if provided
            if !experience.entities.is_empty() {
                existing.experience.entities = experience.entities;
            }

            // Update tags if provided
            if !experience.tags.is_empty() {
                existing.experience.tags = experience.tags;
            }

            // Regenerate embeddings for new content
            let content_hash = Self::sha256_hash(&existing.experience.content);
            if let Some(cached_embedding) = self.content_cache.get(&content_hash) {
                existing.experience.embeddings = Some(cached_embedding.clone());
            } else {
                match self.embedder.encode(&existing.experience.content) {
                    Ok(embedding) => {
                        self.content_cache.insert(content_hash, embedding.clone());
                        existing.experience.embeddings = Some(embedding);
                    }
                    Err(e) => {
                        tracing::warn!("Failed to regenerate embedding during upsert: {}", e);
                    }
                }
            }

            // TEMPORAL EXTRACTION: Re-extract dates when content changes
            let temporal =
                crate::memory::query_parser::extract_temporal_refs(&existing.experience.content);
            existing.experience.temporal_refs.clear();
            for temp_ref in temporal.refs {
                existing
                    .experience
                    .temporal_refs
                    .push(temp_ref.date.to_string());
            }

            // Persist updated memory
            self.long_term_memory.update(&existing)?;

            // Re-index in vector DB with new embeddings
            if let Err(e) = self.retriever.reindex_memory(&existing) {
                tracing::warn!(
                    "Failed to reindex memory {} in vector DB: {}",
                    memory_id.0,
                    e
                );
            }

            // Re-index in BM25 with updated content
            if let Err(e) = self.hybrid_search.index_memory(
                &memory_id,
                &existing.experience.content,
                &existing.experience.tags,
                &existing.experience.entities,
            ) {
                tracing::warn!("Failed to reindex memory {} in BM25: {}", memory_id.0, e);
            }
            if let Err(e) = self.hybrid_search.commit_and_reload() {
                tracing::warn!("Failed to commit/reload BM25 index: {}", e);
            }

            // Update in working/session memory if cached
            {
                let mut working = self.working_memory.write();
                if working.contains(&memory_id) {
                    working.remove(&memory_id)?;
                    working.add_shared(Arc::new(existing.clone()))?;
                }
            }
            {
                let mut session = self.session_memory.write();
                if session.contains(&memory_id) {
                    session.remove(&memory_id)?;
                    session.add_shared(Arc::new(existing.clone()))?;
                }
            }

            tracing::info!(
                external_id = %external_id,
                memory_id = %memory_id.0,
                version = existing.version,
                "Memory upserted (update)"
            );

            Ok((memory_id, true))
        } else {
            // === CREATE PATH ===
            let memory_id = MemoryId(Uuid::new_v4());
            let importance = experience
                .importance_override
                .unwrap_or_else(|| self.calculate_importance(&experience));

            // Generate embeddings if not provided
            if experience.embeddings.is_none() {
                let content_hash = Self::sha256_hash(&experience.content);
                if let Some(cached_embedding) = self.content_cache.get(&content_hash) {
                    experience.embeddings = Some(cached_embedding.clone());
                } else {
                    match self.embedder.encode(&experience.content) {
                        Ok(embedding) => {
                            self.content_cache.insert(content_hash, embedding.clone());
                            experience.embeddings = Some(embedding);
                        }
                        Err(e) => {
                            tracing::warn!("Failed to generate embedding: {}", e);
                        }
                    }
                }
            }

            // TEMPORAL EXTRACTION: Extract dates from content for temporal filtering
            if experience.temporal_refs.is_empty() {
                let temporal =
                    crate::memory::query_parser::extract_temporal_refs(&experience.content);
                for temp_ref in temporal.refs {
                    experience.temporal_refs.push(temp_ref.date.to_string());
                }
            }

            // Create memory with external_id
            let memory = Arc::new(Memory::new_with_external_id(
                memory_id.clone(),
                experience,
                importance,
                external_id.clone(),
                None, // agent_id
                None, // run_id
                None, // actor_id
                None, // created_at
            ));

            // Persist to storage
            self.long_term_memory.store(&memory)?;

            // Log creation
            self.logger.write().log_created(&memory, "working");

            // Add to working memory
            self.working_memory
                .write()
                .add_shared(Arc::clone(&memory))?;

            // Index in vector DB
            if let Err(e) = self.retriever.index_memory(&memory) {
                tracing::warn!("Failed to index memory {} in vector DB: {}", memory.id.0, e);
            }

            // Add entities to knowledge graph with co-occurrence edges
            // PERF: Build entity structs and extract co-occurrences OUTSIDE the lock
            if let Some(graph) = &self.graph_memory {
                let now = chrono::Utc::now();

                // Phase 1: Build entity structs with proper labels from NER
                let ner_lookup = build_ner_lookup(&memory.experience.ner_entities);

                // Batch-encode entity names for concept-level dedup
                let entity_names: Vec<&str> = memory
                    .experience
                    .entities
                    .iter()
                    .map(|s| s.as_str())
                    .collect();
                let entity_embeddings: Vec<Option<Vec<f32>>> = if entity_names.is_empty() {
                    Vec::new()
                } else {
                    match self.embedder.encode_batch(&entity_names) {
                        Ok(embs) => embs.into_iter().map(Some).collect(),
                        Err(e) => {
                            tracing::debug!(
                                error = %e,
                                "Entity name embedding failed, concept merge disabled for this batch"
                            );
                            vec![None; entity_names.len()]
                        }
                    }
                };

                let entities_to_add: Vec<crate::graph_memory::EntityNode> = memory
                    .experience
                    .entities
                    .iter()
                    .zip(entity_embeddings.into_iter())
                    .map(|(entity_name, embedding)| {
                        let (label, salience) = resolve_entity_label(entity_name, &ner_lookup);
                        crate::graph_memory::EntityNode {
                            uuid: Uuid::new_v4(),
                            name: entity_name.clone(),
                            labels: vec![label],
                            created_at: now,
                            last_seen_at: now,
                            mention_count: 1,
                            summary: String::new(),
                            attributes: std::collections::HashMap::new(),
                            name_embedding: embedding,
                            salience,
                            is_proper_noun: entity_name
                                .chars()
                                .next()
                                .map(|c| c.is_uppercase())
                                .unwrap_or(false),
                            selectivity: None,
                        }
                    })
                    .collect();

                // Phase 2: Use pre-extracted co-occurrence pairs or extract fresh
                let cooccurrence_pairs = if !memory.experience.cooccurrence_pairs.is_empty() {
                    memory.experience.cooccurrence_pairs.clone()
                } else {
                    let entity_extractor = crate::graph_memory::EntityExtractor::new();
                    entity_extractor.extract_cooccurrence_pairs(&memory.experience.content)
                };

                let edge_context = format!("Co-occurred in memory {}", memory.id.0);

                // Phase 3: Acquire read lock for graph insertions (GraphMemory is internally thread-safe)
                let graph_guard = graph.read();

                for entity in entities_to_add {
                    if let Err(e) = graph_guard.add_entity(entity.clone()) {
                        tracing::debug!("Failed to add entity '{}' to graph: {}", entity.name, e);
                    }
                }

                // Semantic edge weighting
                let l1_base_weight = crate::graph_memory::EdgeTier::L1Working.initial_weight();
                for (entity1, entity2) in cooccurrence_pairs {
                    if let (Ok(Some(e1)), Ok(Some(e2))) = (
                        graph_guard.find_entity_by_name(&entity1),
                        graph_guard.find_entity_by_name(&entity2),
                    ) {
                        let entity_confidence = Some((e1.salience + e2.salience) / 2.0);

                        let semantic_weight = match (&e1.name_embedding, &e2.name_embedding) {
                            (Some(emb1), Some(emb2)) => {
                                let sim = crate::similarity::cosine_similarity(emb1, emb2).max(0.0);
                                EDGE_SEMANTIC_WEIGHT_FLOOR
                                    + (1.0 - EDGE_SEMANTIC_WEIGHT_FLOOR) * sim
                            }
                            _ => 1.0,
                        };

                        let edge = crate::graph_memory::RelationshipEdge {
                            uuid: Uuid::new_v4(),
                            from_entity: e1.uuid,
                            to_entity: e2.uuid,
                            relation_type: crate::graph_memory::RelationType::CoOccurs,
                            strength: l1_base_weight * semantic_weight,
                            created_at: now,
                            valid_at: now,
                            invalidated_at: None,
                            source_episode_id: Some(memory.id.0),
                            context: edge_context.clone(),
                            last_activated: now,
                            activation_count: 1,
                            ltp_status: crate::graph_memory::LtpStatus::None,
                            tier: crate::graph_memory::EdgeTier::L1Working,
                            activation_timestamps: None,
                            entity_confidence,
                            forman_curvature: None,
                            endpoint_selectivity: None,
                        };

                        if let Err(e) = graph_guard.add_relationship(edge) {
                            tracing::trace!(
                                "Failed to add co-occurrence edge {}<->{}: {}",
                                entity1,
                                entity2,
                                e
                            );
                        }
                    }
                }
            }

            // Index in BM25 for hybrid search
            if let Err(e) = self.hybrid_search.index_memory(
                &memory.id,
                &memory.experience.content,
                &memory.experience.tags,
                &memory.experience.entities,
            ) {
                tracing::warn!("Failed to index memory {} in BM25: {}", memory.id.0, e);
            }
            if let Err(e) = self.hybrid_search.commit_and_reload() {
                tracing::warn!("Failed to commit/reload BM25 index: {}", e);
            }

            // Add to session if important
            if importance > self.config.importance_threshold {
                self.session_memory
                    .write()
                    .add_shared(Arc::clone(&memory))?;
            }

            // Update stats
            self.stats.write().total_memories += 1;

            tracing::info!(
                external_id = %external_id,
                memory_id = %memory_id.0,
                "Memory upserted (create)"
            );

            Ok((memory_id, false))
        }
    }

    /// Get the history of a memory (audit trail of changes)
    ///
    /// Returns the full revision history for memories with external linking.
    /// Returns empty vec for regular (non-mutable) memories.
    pub fn get_memory_history(&self, memory_id: &MemoryId) -> Result<Vec<MemoryRevision>> {
        let memory = self.long_term_memory.get(memory_id)?;
        Ok(memory.history.clone())
    }

    /// Find a memory by external ID
    ///
    /// Used to check if a memory already exists for an external entity
    pub fn find_by_external_id(&self, external_id: &str) -> Result<Option<Memory>> {
        self.long_term_memory.find_by_external_id(external_id)
    }

    /// Run periodic maintenance (consolidation, activation decay, graph maintenance)
    ///
    /// Call this periodically (e.g., every 5 minutes) to:
    /// 1. Promote memories between tiers based on thresholds
    /// 2. Decay activation levels on all memories
    /// 3. Run graph maintenance (prune weak edges)
    ///
    /// `is_heavy`: when true, runs expensive operations (fact extraction, auto-repair)
    /// that require full RocksDB scans. Light cycles only touch in-memory data.
    ///
    /// Returns the number of memories processed for activation decay.
    /// Also records consolidation events for introspection.
    pub fn run_maintenance(
        &self,
        decay_factor: f32,
        user_id: &str,
        is_heavy: bool,
    ) -> Result<MaintenanceResult> {
        let start_time = std::time::Instant::now();
        let now = chrono::Utc::now();

        // 1. Consolidation: promote memories between tiers
        self.consolidate_if_needed()?;

        // 2. Decay activation on all in-memory memories (working + session)
        let mut decayed_count = 0;
        let mut at_risk_count = 0;
        const AT_RISK_THRESHOLD: f32 = 0.2; // Memories below this are at risk of being forgotten

        // Decay working memory activations with event tracking
        // Emotional modulation: high-arousal memories decay slower (amygdala coupling)
        {
            let working = self.working_memory.read();
            for memory in working.all_memories() {
                let activation_before = memory.activation();
                let arousal = memory
                    .experience
                    .context
                    .as_ref()
                    .map(|c| c.emotional.arousal)
                    .unwrap_or(0.0);
                let emotional_factor =
                    1.0 - (arousal * crate::constants::EMOTIONAL_DECAY_MODULATION);
                memory.decay_activation(decay_factor * emotional_factor);
                let activation_after = memory.activation();
                decayed_count += 1;

                // Only record event if there was actual decay
                if activation_before != activation_after {
                    let at_risk = activation_after < AT_RISK_THRESHOLD;
                    if at_risk {
                        at_risk_count += 1;
                    }

                    // Record decay event
                    self.record_consolidation_event(ConsolidationEvent::MemoryDecayed {
                        memory_id: memory.id.0.to_string(),
                        content_preview: memory.experience.content.chars().take(50).collect(),
                        activation_before,
                        activation_after,
                        at_risk,
                        timestamp: now,
                    });
                }
            }
        }

        // Decay session memory activations with event tracking
        // Emotional modulation: high-arousal memories decay slower (amygdala coupling)
        {
            let session = self.session_memory.read();
            for memory in session.all_memories() {
                let activation_before = memory.activation();
                let arousal = memory
                    .experience
                    .context
                    .as_ref()
                    .map(|c| c.emotional.arousal)
                    .unwrap_or(0.0);
                let emotional_factor =
                    1.0 - (arousal * crate::constants::EMOTIONAL_DECAY_MODULATION);
                memory.decay_activation(decay_factor * emotional_factor);
                let activation_after = memory.activation();
                decayed_count += 1;

                // Only record event if there was actual decay
                if activation_before != activation_after {
                    let at_risk = activation_after < AT_RISK_THRESHOLD;
                    if at_risk {
                        at_risk_count += 1;
                    }

                    // Record decay event
                    self.record_consolidation_event(ConsolidationEvent::MemoryDecayed {
                        memory_id: memory.id.0.to_string(),
                        content_preview: memory.experience.content.chars().take(50).collect(),
                        activation_before,
                        activation_after,
                        at_risk,
                        timestamp: now,
                    });
                }
            }
        }

        // 2.5 Potentiation: boost ALL memories based on access count (Hebbian LTP)
        // This implements "neurons that fire together wire together" - memories
        // that are accessed frequently get importance boosts during maintenance
        let mut potentiated_count = 0;
        {
            // Potentiate working memory
            let working = self.working_memory.read();
            for memory in working.all_memories() {
                // Only boost if below saturation threshold (0.95) to prevent
                // all frequently-accessed memories converging to importance=1.0
                if memory.access_count() >= POTENTIATION_ACCESS_THRESHOLD
                    && memory.importance() < 0.95
                {
                    let activation_before = memory.importance();
                    memory.boost_importance(POTENTIATION_MAINTENANCE_BOOST);
                    potentiated_count += 1;

                    self.record_consolidation_event(ConsolidationEvent::MemoryStrengthened {
                        memory_id: memory.id.0.to_string(),
                        content_preview: memory.experience.content.chars().take(50).collect(),
                        activation_before,
                        activation_after: memory.importance(),
                        reason: StrengtheningReason::MaintenancePotentiation,
                        timestamp: now,
                    });
                }
            }
        }
        {
            // Potentiate session memory
            let session = self.session_memory.read();
            for memory in session.all_memories() {
                if memory.access_count() >= POTENTIATION_ACCESS_THRESHOLD
                    && memory.importance() < 0.95
                {
                    let activation_before = memory.importance();
                    memory.boost_importance(POTENTIATION_MAINTENANCE_BOOST);
                    potentiated_count += 1;

                    self.record_consolidation_event(ConsolidationEvent::MemoryStrengthened {
                        memory_id: memory.id.0.to_string(),
                        content_preview: memory.experience.content.chars().take(50).collect(),
                        activation_before,
                        activation_after: memory.importance(),
                        reason: StrengtheningReason::MaintenancePotentiation,
                        timestamp: now,
                    });
                }
            }
        }

        if potentiated_count > 0 {
            tracing::debug!(
                "Potentiated {} memories during maintenance (access >= {})",
                potentiated_count,
                POTENTIATION_ACCESS_THRESHOLD
            );
        }

        // 3. Graph maintenance moved to state.rs run_maintenance_all_users()
        // This fixes the double-decay bug: apply_decay() was called both here
        // (via graph_maintenance()) and in state.rs. Now only state.rs calls it,
        // and the result is used for orphan detection (Direction 2 coupling).

        // 3.5. Temporal fact decay: decay/delete stale facts (heavy only)
        // Scans all facts via RocksDB iterator — deferred to heavy cycles
        let (facts_decayed, facts_deleted) = if is_heavy {
            self.decay_facts_for_all_users().unwrap_or((0, 0))
        } else {
            (0, 0)
        };
        if facts_decayed > 0 || facts_deleted > 0 {
            tracing::debug!(
                "Temporal fact maintenance: {} decayed, {} deleted",
                facts_decayed,
                facts_deleted
            );
        }

        // 3.7. Heavy cycle: load all memories once for both fact extraction and replay.
        // This avoids two separate RocksDB full scans on the same cycle.
        let all_memories_for_heavy: Vec<SharedMemory> = if is_heavy {
            self.get_all_memories().unwrap_or_default()
        } else {
            Vec::new()
        };

        // 3.8. Fact extraction: consolidate episodic memories into semantic facts
        // HEAVY ONLY: requires ONNX inference for embedding new facts.
        // The dirty flag (fact_extraction_needed) is only checked on heavy cycles;
        // it stays set across light cycles until the next heavy cycle processes it.
        let mut facts_extracted_count = 0;
        let mut facts_reinforced_count = 0;
        if is_heavy
            && self
                .fact_extraction_needed
                .swap(false, std::sync::atomic::Ordering::Relaxed)
        {
            let all_memories = &all_memories_for_heavy;
            if !all_memories.is_empty() {
                // Incremental: only process memories created since last extraction watermark.
                // First run (watermark=0) processes everything; subsequent runs only new memories.
                // Lazy init: if watermark is 0 (startup sentinel), load persisted value
                // or derive from the latest fact's created_at timestamp.
                let mut watermark_millis = self
                    .fact_extraction_watermark
                    .load(std::sync::atomic::Ordering::Relaxed);
                if watermark_millis == 0 {
                    watermark_millis = self
                        .long_term_memory
                        .get_fact_watermark(user_id)
                        .or_else(|| self.fact_store.latest_fact_created_at(user_id))
                        .unwrap_or(0);
                    if watermark_millis > 0 {
                        self.fact_extraction_watermark
                            .store(watermark_millis, std::sync::atomic::Ordering::Relaxed);
                    }
                }
                let watermark_dt = chrono::DateTime::from_timestamp_millis(watermark_millis)
                    .unwrap_or(chrono::DateTime::<chrono::Utc>::MIN_UTC);

                let memories: Vec<Memory> = all_memories
                    .iter()
                    .filter(|m| m.created_at > watermark_dt)
                    .map(|arc_mem| arc_mem.as_ref().clone())
                    .collect();

                tracing::info!(
                    total_memories = all_memories.len(),
                    new_since_watermark = memories.len(),
                    watermark = %watermark_dt.format("%Y-%m-%dT%H:%M:%S"),
                    "Incremental fact extraction"
                );

                let consolidator = compression::SemanticConsolidator::new();
                let consolidation_result = consolidator.consolidate(&memories);

                if !consolidation_result.new_facts.is_empty() {
                    // Batch-encode all new fact texts for hybrid dedup
                    let fact_texts: Vec<&str> = consolidation_result
                        .new_facts
                        .iter()
                        .map(|f| f.fact.as_str())
                        .collect();
                    let fact_embeddings: Vec<Option<Vec<f32>>> = match self
                        .embedder
                        .encode_batch(&fact_texts)
                    {
                        Ok(embs) => embs.into_iter().map(Some).collect(),
                        Err(e) => {
                            tracing::debug!(
                                error = %e,
                                "Fact embedding batch failed, falling back to Jaccard-only dedup"
                            );
                            vec![None; fact_texts.len()]
                        }
                    };

                    let mut truly_new: Vec<(SemanticFact, Option<Vec<f32>>)> = Vec::new();

                    for (fact, embedding) in consolidation_result
                        .new_facts
                        .iter()
                        .zip(fact_embeddings.into_iter())
                    {
                        // Hybrid dedup: embedding cosine + entity gate + polarity + Jaccard floor
                        match self.fact_store.find_similar(
                            user_id,
                            &fact.fact,
                            &fact.related_entities,
                            embedding.as_deref(),
                        ) {
                            Ok(Some(mut existing)) => {
                                // Reinforce the existing fact
                                existing.support_count += 1;
                                existing.last_reinforced = now;
                                let confidence_before = existing.confidence;
                                let boost = 0.1 * (1.0 - existing.confidence);
                                existing.confidence = (existing.confidence + boost).min(1.0);

                                // Extend source memories and related entities
                                for src in &fact.source_memories {
                                    if !existing.source_memories.contains(src) {
                                        existing.source_memories.push(src.clone());
                                    }
                                }
                                for entity in &fact.related_entities {
                                    if !existing.related_entities.contains(entity) {
                                        existing.related_entities.push(entity.clone());
                                    }
                                }

                                if let Err(e) = self.fact_store.update(user_id, &existing) {
                                    tracing::debug!("Failed to reinforce fact: {e}");
                                } else {
                                    // Update existing fact's embedding with latest encoding
                                    if let Some(ref emb) = embedding {
                                        let _ = self.fact_store.store_embedding(
                                            user_id,
                                            &existing.id,
                                            emb,
                                        );
                                    }
                                    facts_reinforced_count += 1;
                                    self.record_consolidation_event_for_user(
                                        user_id,
                                        ConsolidationEvent::FactReinforced {
                                            fact_id: existing.id.clone(),
                                            fact_content: existing.fact.clone(),
                                            confidence_before,
                                            confidence_after: existing.confidence,
                                            new_support_count: existing.support_count,
                                            timestamp: now,
                                        },
                                    );
                                }
                            }
                            _ => {
                                truly_new.push((fact.clone(), embedding));
                            }
                        }
                    }

                    // Store new facts
                    if !truly_new.is_empty() {
                        let facts_only: Vec<SemanticFact> =
                            truly_new.iter().map(|(f, _)| f.clone()).collect();
                        match self.fact_store.store_batch(user_id, &facts_only) {
                            Ok(stored) => {
                                facts_extracted_count = stored;
                                // Store embeddings for newly persisted facts
                                for (fact, embedding) in &truly_new {
                                    if let Some(emb) = embedding {
                                        let _ =
                                            self.fact_store.store_embedding(user_id, &fact.id, emb);
                                    }
                                    self.record_consolidation_event_for_user(
                                        user_id,
                                        ConsolidationEvent::FactExtracted {
                                            fact_id: fact.id.clone(),
                                            fact_content: fact.fact.clone(),
                                            confidence: fact.confidence,
                                            fact_type: format!("{:?}", fact.fact_type),
                                            source_memory_count: fact.source_memories.len(),
                                            timestamp: now,
                                        },
                                    );
                                }
                            }
                            Err(e) => {
                                tracing::warn!("Failed to store extracted facts: {e}");
                            }
                        }

                        // Connect newly extracted facts to the knowledge graph
                        self.connect_facts_to_graph(&facts_only);
                    }

                    if facts_extracted_count > 0 || facts_reinforced_count > 0 {
                        tracing::debug!(
                            extracted = facts_extracted_count,
                            reinforced = facts_reinforced_count,
                            "Fact consolidation during maintenance"
                        );
                    }
                }

                // Advance watermark to the LAST memory's created_at timestamp,
                // NOT to now(). Using now() would skip memories created during the
                // (potentially slow) fact extraction cycle — they'd have created_at
                // < now() and never be processed for facts.
                if !memories.is_empty() {
                    let new_watermark = memories
                        .iter()
                        .map(|m| m.created_at.timestamp_millis())
                        .max()
                        .unwrap_or_else(|| chrono::Utc::now().timestamp_millis());
                    self.fact_extraction_watermark
                        .store(new_watermark, std::sync::atomic::Ordering::Relaxed);
                    self.long_term_memory
                        .set_fact_watermark(user_id, new_watermark);
                }
            }
        } else {
            tracing::debug!("Fact extraction skipped: no new memories since last cycle");
        }

        // 4. SHO-105 + PIPE-2: Memory replay cycle (consolidation during heavy cycles)
        //
        // HEAVY ONLY: Replay draws candidates from ALL memory tiers (including long-term
        // via the shared all_memories_for_heavy loaded above). Light cycles skip replay
        // entirely — analogous to "consolidation during deep sleep, not during waking."
        //
        // Pattern detection still runs to record triggers, but actual replay execution
        // only happens on heavy cycles where we have the full memory corpus.
        let mut replay_result = replay::ReplayCycleResult::default();
        {
            // PIPE-2: Check for pattern-triggered replay first
            let pattern_result = self.pattern_detector.write().detect_patterns();
            let has_pattern_triggers = !pattern_result.triggers.is_empty();

            // Log pattern detection results
            if has_pattern_triggers {
                tracing::debug!(
                    "Pattern detection: {} entity, {} semantic, {} temporal, {} salience, {} behavioral triggers",
                    pattern_result.entity_patterns_found,
                    pattern_result.semantic_clusters_found,
                    pattern_result.temporal_clusters_found,
                    pattern_result.salience_spikes_found,
                    pattern_result.behavioral_changes_found
                );

                // Record pattern-triggered replay events
                for trigger in &pattern_result.triggers {
                    self.record_consolidation_event(
                        introspection::ConsolidationEvent::PatternTriggeredReplay {
                            trigger_type: trigger.trigger_type_name().to_string(),
                            memory_ids: trigger.memory_ids(),
                            pattern_confidence: match trigger {
                                pattern_detection::ReplayTrigger::EntityCoOccurrence {
                                    confidence,
                                    ..
                                } => *confidence,
                                pattern_detection::ReplayTrigger::SemanticCluster {
                                    avg_similarity,
                                    ..
                                } => *avg_similarity,
                                pattern_detection::ReplayTrigger::SalienceSpike {
                                    importance,
                                    ..
                                } => *importance,
                                _ => 0.7, // Default confidence for other triggers
                            },
                            trigger_description: trigger.description(),
                            timestamp: now,
                        },
                    );
                }
            }

            // Replay only on heavy cycles — uses shared all_memories_for_heavy
            let timer_should_replay = self.replay_manager.read().should_replay();
            let should_replay = is_heavy && (has_pattern_triggers || timer_should_replay);

            if should_replay && !all_memories_for_heavy.is_empty() {
                // Build replay candidates from ALL memory tiers (not just working+session)
                let graph_ref = self.graph_memory.clone();
                let candidates_data: Vec<_> = all_memories_for_heavy
                    .iter()
                    .map(|m| {
                        // Fetch actual connections from GraphMemory
                        let connections: Vec<String> = if let Some(ref graph) = graph_ref {
                            let graph_guard = graph.read();
                            graph_guard
                                .find_memory_associations(&m.id.0, 10)
                                .unwrap_or_default()
                                .into_iter()
                                .map(|(uuid, _)| uuid.to_string())
                                .collect()
                        } else {
                            Vec::new()
                        };
                        let arousal = m
                            .experience
                            .context
                            .as_ref()
                            .map(|c| c.emotional.arousal)
                            .unwrap_or(0.3);
                        (
                            m.id.0.to_string(),
                            m.importance(),
                            arousal,
                            m.created_at,
                            connections,
                            m.experience.content.chars().take(50).collect::<String>(),
                        )
                    })
                    .collect();

                // Identify and execute replay
                let candidates = self
                    .replay_manager
                    .read()
                    .identify_replay_candidates(&candidates_data);

                if !candidates.is_empty() {
                    let (memory_boosts, edge_boosts, events) =
                        self.replay_manager.write().execute_replay(&candidates);

                    replay_result.memories_replayed = candidates.len();
                    replay_result.edges_strengthened = edge_boosts.len();
                    replay_result.total_priority_score =
                        candidates.iter().map(|c| c.priority_score).sum();

                    // Collect replayed memory IDs for entity-entity edge strengthening
                    replay_result.replay_memory_ids =
                        candidates.iter().map(|c| c.memory_id.clone()).collect();

                    // Apply memory boosts
                    for (mem_id_str, boost) in &memory_boosts {
                        if let Ok(mem_id) = uuid::Uuid::parse_str(mem_id_str) {
                            if let Ok(memory) = self.long_term_memory.get(&MemoryId(mem_id)) {
                                memory.boost_importance(*boost);
                                if let Err(e) = self.long_term_memory.update(&memory) {
                                    tracing::debug!("Failed to persist replay boost: {e}");
                                }
                            }
                        }
                    }

                    // Collect edge boosts to return - will be applied via GraphMemory at API layer
                    replay_result.edge_boosts = edge_boosts;
                    if !replay_result.edge_boosts.is_empty() {
                        tracing::debug!(
                            "Replay produced {} edge boosts (to be applied via GraphMemory)",
                            replay_result.edge_boosts.len()
                        );
                    }

                    // Record events
                    for event in events {
                        self.record_consolidation_event(event);
                    }

                    // Record replay cycle completion
                    self.record_consolidation_event(ConsolidationEvent::ReplayCycleCompleted {
                        memories_replayed: replay_result.memories_replayed,
                        edges_strengthened: replay_result.edges_strengthened,
                        total_priority_score: replay_result.total_priority_score,
                        duration_ms: start_time.elapsed().as_millis() as u64,
                        timestamp: now,
                    });

                    tracing::debug!(
                        "Replay cycle complete: {} memories replayed, {} edges strengthened",
                        replay_result.memories_replayed,
                        replay_result.edges_strengthened
                    );
                }
            }
        }

        // 4.5. PIPE-2: Cleanup old patterns to prevent unbounded memory growth
        // Removes patterns older than 24 hours
        self.pattern_detector.write().cleanup();

        // 5. Auto-repair index integrity and compact if needed (heavy only)
        // repair_vector_index() does a full RocksDB scan + ONNX inference per orphan
        if is_heavy {
            self.auto_repair_and_compact();
        }

        let duration_ms = start_time.elapsed().as_millis() as u64;

        // Record maintenance cycle completion event
        self.record_consolidation_event(ConsolidationEvent::MaintenanceCycleCompleted {
            memories_processed: decayed_count,
            memories_decayed: decayed_count, // All memories get decay applied
            edges_pruned: 0,                 // Graph maintenance doesn't report this yet
            duration_ms,
            timestamp: now,
        });

        tracing::debug!(
            "Maintenance complete: {} memories decayed (factor={}), {} at risk, {} replayed, took {}ms",
            decayed_count,
            decay_factor,
            at_risk_count,
            replay_result.memories_replayed,
            duration_ms
        );

        Ok(MaintenanceResult {
            decayed_count,
            edge_boosts: replay_result.edge_boosts,
            replay_memory_ids: replay_result.replay_memory_ids,
            memories_replayed: replay_result.memories_replayed,
            total_priority_score: replay_result.total_priority_score,
            facts_extracted: facts_extracted_count,
            facts_reinforced: facts_reinforced_count,
        })
    }

    // =========================================================================
    // CONSOLIDATION INTROSPECTION API
    // =========================================================================

    /// Get a consolidation report for a time period
    ///
    /// Shows what the memory system has been learning:
    /// - Which memories strengthened or decayed
    /// - What associations formed or were pruned
    /// - What facts were extracted or reinforced
    ///
    /// # Arguments
    /// * `since` - Start of the time period
    /// * `until` - End of the time period (default: now)
    pub fn get_consolidation_report(
        &self,
        since: chrono::DateTime<chrono::Utc>,
        until: Option<chrono::DateTime<chrono::Utc>>,
    ) -> ConsolidationReport {
        let until = until.unwrap_or_else(chrono::Utc::now);
        let events = self.consolidation_events.read();
        events.generate_report(since, until)
    }

    /// Get a consolidation report for a user using persisted history
    ///
    /// Unlike `get_consolidation_report`, this method uses persisted learning history
    /// and can generate reports spanning across restarts. It combines:
    /// - Persisted significant events from learning_history (survives restarts)
    /// - Ephemeral events from the event buffer (current session)
    ///
    /// Use this when you need historical reports beyond the current session.
    pub fn get_consolidation_report_for_user(
        &self,
        user_id: &str,
        since: chrono::DateTime<chrono::Utc>,
        until: Option<chrono::DateTime<chrono::Utc>>,
    ) -> Result<ConsolidationReport> {
        let until = until.unwrap_or_else(chrono::Utc::now);

        // Get persisted significant events from learning history
        let persisted_events = self
            .learning_history
            .events_in_range(user_id, since, until)?;

        // Get ephemeral events from the buffer
        let ephemeral_events = {
            let events = self.consolidation_events.read();
            events.events_since(since)
        };

        // Combine events, deduplicating by (timestamp, event_type) to avoid
        // dropping distinct events that share a nanosecond timestamp.
        let mut all_events: Vec<ConsolidationEvent> = Vec::new();
        let mut seen_keys: std::collections::HashSet<(
            i64,
            std::mem::Discriminant<ConsolidationEvent>,
        )> = std::collections::HashSet::new();

        // Add persisted events first (these are significant events that survived restart)
        for stored in &persisted_events {
            let ts = stored.event.timestamp().timestamp_nanos_opt().unwrap_or(0);
            let key = (ts, std::mem::discriminant(&stored.event));
            if seen_keys.insert(key) {
                all_events.push(stored.event.clone());
            }
        }

        // Add ephemeral events that aren't already included
        let until_nanos = until.timestamp_nanos_opt().unwrap_or(i64::MAX);
        for event in ephemeral_events {
            let ts = event.timestamp().timestamp_nanos_opt().unwrap_or(0);
            let key = (ts, std::mem::discriminant(&event));
            if ts <= until_nanos && seen_keys.insert(key) {
                all_events.push(event);
            }
        }

        // Sort by timestamp
        all_events.sort_by(|a, b| a.timestamp().cmp(&b.timestamp()));

        // Generate report from combined events
        let report =
            ConsolidationEventBuffer::generate_report_from_events(&all_events, since, until);

        Ok(report)
    }

    /// Get all consolidation events since a timestamp
    ///
    /// Returns raw events for detailed analysis
    pub fn get_consolidation_events_since(
        &self,
        since: chrono::DateTime<chrono::Utc>,
    ) -> Vec<ConsolidationEvent> {
        let events = self.consolidation_events.read();
        events.events_since(since)
    }

    /// Get all consolidation events in the buffer
    pub fn get_all_consolidation_events(&self) -> Vec<ConsolidationEvent> {
        let events = self.consolidation_events.read();
        events.all_events()
    }

    /// Record a consolidation event
    ///
    /// Used internally by the memory system to log learning events.
    /// Also available for external callers that want to track custom events.
    pub fn record_consolidation_event(&self, event: ConsolidationEvent) {
        let mut events = self.consolidation_events.write();
        events.push(event);
    }

    /// Record a consolidation event for a specific user
    ///
    /// This method both:
    /// 1. Pushes to the ephemeral event buffer (for real-time introspection)
    /// 2. Persists significant events to learning_history (for retrieval boosting)
    ///
    /// Use this instead of `record_consolidation_event` when you have a user_id.
    pub fn record_consolidation_event_for_user(&self, user_id: &str, event: ConsolidationEvent) {
        // Always push to ephemeral buffer
        {
            let mut events = self.consolidation_events.write();
            events.push(event.clone());
        }

        // Persist significant events to learning history
        if event.is_significant() {
            if let Err(e) = self.learning_history.record(user_id, &event) {
                tracing::warn!(
                    user_id = %user_id,
                    event_type = ?std::mem::discriminant(&event),
                    error = %e,
                    "Failed to persist learning event"
                );
            }
        }
    }

    /// Clear all consolidation events
    pub fn clear_consolidation_events(&self) {
        let mut events = self.consolidation_events.write();
        events.clear();
    }

    /// Get the number of consolidation events in the buffer
    pub fn consolidation_event_count(&self) -> usize {
        let events = self.consolidation_events.read();
        events.len()
    }

    // =========================================================================
    // SEMANTIC FACT OPERATIONS (SHO-f0e7)
    // Distilled knowledge extracted from episodic memories
    // =========================================================================

    /// Distill semantic facts from episodic memories
    ///
    /// Runs the consolidation process to extract durable knowledge:
    /// 1. Find patterns appearing in multiple memories
    /// 2. Create or reinforce semantic facts
    /// 3. Store facts in the fact store
    ///
    /// # Arguments
    /// * `user_id` - User whose memories to consolidate
    /// * `min_support` - Minimum memories needed to form a fact (default: 3)
    /// * `min_age_days` - Minimum age of memories to consider (default: 7)
    ///
    /// # Returns
    /// ConsolidationResult with stats and newly extracted facts
    pub fn distill_facts(
        &self,
        user_id: &str,
        min_support: usize,
        min_age_days: i64,
    ) -> Result<ConsolidationResult> {
        // Get all memories for consolidation
        let all_memories = self.get_all_memories()?;

        // Incremental: only process memories created since last extraction watermark
        let mut watermark_millis = self
            .fact_extraction_watermark
            .load(std::sync::atomic::Ordering::Relaxed);
        if watermark_millis == 0 {
            watermark_millis = self
                .long_term_memory
                .get_fact_watermark(user_id)
                .or_else(|| self.fact_store.latest_fact_created_at(user_id))
                .unwrap_or(0);
            if watermark_millis > 0 {
                self.fact_extraction_watermark
                    .store(watermark_millis, std::sync::atomic::Ordering::Relaxed);
            }
        }
        let watermark_dt = chrono::DateTime::from_timestamp_millis(watermark_millis)
            .unwrap_or(chrono::DateTime::<chrono::Utc>::MIN_UTC);

        let memories: Vec<Memory> = all_memories
            .iter()
            .filter(|m| m.created_at > watermark_dt)
            .map(|arc_mem| arc_mem.as_ref().clone())
            .collect();

        tracing::info!(
            total_memories = all_memories.len(),
            new_since_watermark = memories.len(),
            watermark = %watermark_dt.format("%Y-%m-%dT%H:%M:%S"),
            "Incremental fact extraction (on-demand)"
        );

        // Create consolidator with custom thresholds
        let consolidator =
            compression::SemanticConsolidator::with_thresholds(min_support, min_age_days);

        // Run consolidation
        let result = consolidator.consolidate(&memories);

        // Store extracted facts
        if !result.new_facts.is_empty() {
            let stored = self.fact_store.store_batch(user_id, &result.new_facts)?;
            tracing::info!(
                user_id = %user_id,
                facts_extracted = result.facts_extracted,
                facts_stored = stored,
                "Semantic distillation complete"
            );

            // Batch-encode and store embeddings for distilled facts
            let texts: Vec<&str> = result.new_facts.iter().map(|f| f.fact.as_str()).collect();
            if let Ok(batch_embs) = self.embedder.encode_batch(&texts) {
                for (fact, emb) in result.new_facts.iter().zip(batch_embs.iter()) {
                    let _ = self.fact_store.store_embedding(user_id, &fact.id, emb);
                }
            }

            // Record consolidation event for each fact (persists significant events)
            for fact in &result.new_facts {
                self.record_consolidation_event_for_user(
                    user_id,
                    ConsolidationEvent::FactExtracted {
                        fact_id: fact.id.clone(),
                        fact_content: fact.fact.clone(),
                        confidence: fact.confidence,
                        fact_type: format!("{:?}", fact.fact_type),
                        source_memory_count: fact.source_memories.len(),
                        timestamp: chrono::Utc::now(),
                    },
                );
            }
        }

        // Advance watermark after successful extraction
        if !memories.is_empty() {
            let new_watermark = chrono::Utc::now().timestamp_millis();
            self.fact_extraction_watermark
                .store(new_watermark, std::sync::atomic::Ordering::Relaxed);
            self.long_term_memory
                .set_fact_watermark(user_id, new_watermark);
        }

        Ok(result)
    }

    /// Get semantic facts for a user
    ///
    /// # Arguments
    /// * `user_id` - User whose facts to retrieve
    /// * `limit` - Maximum number of facts to return
    pub fn get_facts(&self, user_id: &str, limit: usize) -> Result<Vec<SemanticFact>> {
        self.fact_store.list(user_id, limit)
    }

    /// Get facts related to a specific entity
    ///
    /// # Arguments
    /// * `user_id` - User whose facts to search
    /// * `entity` - Entity to search for (e.g., "authentication", "JWT")
    /// * `limit` - Maximum number of facts to return
    pub fn get_facts_by_entity(
        &self,
        user_id: &str,
        entity: &str,
        limit: usize,
    ) -> Result<Vec<SemanticFact>> {
        self.fact_store.find_by_entity(user_id, entity, limit)
    }

    /// Get facts of a specific type
    ///
    /// # Arguments
    /// * `user_id` - User whose facts to search
    /// * `fact_type` - Type of fact (Preference, Procedure, Definition, etc.)
    /// * `limit` - Maximum number of facts to return
    pub fn get_facts_by_type(
        &self,
        user_id: &str,
        fact_type: FactType,
        limit: usize,
    ) -> Result<Vec<SemanticFact>> {
        self.fact_store.find_by_type(user_id, fact_type, limit)
    }

    /// Search facts by keyword
    ///
    /// # Arguments
    /// * `user_id` - User whose facts to search
    /// * `query` - Search query
    /// * `limit` - Maximum number of facts to return
    pub fn search_facts(
        &self,
        user_id: &str,
        query: &str,
        limit: usize,
    ) -> Result<Vec<SemanticFact>> {
        self.fact_store.search(user_id, query, limit)
    }

    /// Get statistics about stored facts
    pub fn get_fact_stats(&self, user_id: &str) -> Result<facts::FactStats> {
        self.fact_store.stats(user_id)
    }

    /// Get facts associated with graph entity names.
    ///
    /// Bridges graph traversal → fact retrieval: when spreading activation discovers
    /// entity nodes, this method returns the semantic facts linked to those entities.
    /// Results are deduplicated and sorted by confidence (highest first).
    pub fn get_facts_for_graph_entities(
        &self,
        user_id: &str,
        entity_names: &[String],
        limit_per_entity: usize,
    ) -> Result<Vec<SemanticFact>> {
        let mut seen_ids = std::collections::HashSet::new();
        let mut results = Vec::new();

        for name in entity_names {
            let facts = self
                .fact_store
                .find_by_entity(user_id, name, limit_per_entity)?;
            for fact in facts {
                if seen_ids.insert(fact.id.clone()) {
                    results.push(fact);
                }
            }
        }

        results.sort_by(|a, b| b.confidence.total_cmp(&a.confidence));
        Ok(results)
    }

    /// Reinforce a fact with new supporting evidence
    ///
    /// Called when a new memory supports an existing fact.
    /// Increments support_count and boosts confidence.
    pub fn reinforce_fact(
        &self,
        user_id: &str,
        fact_id: &str,
        memory_id: &MemoryId,
    ) -> Result<bool> {
        if let Some(mut fact) = self.fact_store.get(user_id, fact_id)? {
            // Track confidence before change for event
            let confidence_before = fact.confidence;

            // Increment support
            fact.support_count += 1;
            fact.last_reinforced = chrono::Utc::now();

            // Boost confidence with diminishing returns
            let boost = 0.1 * (1.0 - fact.confidence);
            fact.confidence = (fact.confidence + boost).min(1.0);

            // Add source if not already present
            if !fact.source_memories.contains(memory_id) {
                fact.source_memories.push(memory_id.clone());
            }

            // Update in store
            self.fact_store.update(user_id, &fact)?;

            // Record reinforcement event (persists significant events)
            self.record_consolidation_event_for_user(
                user_id,
                ConsolidationEvent::FactReinforced {
                    fact_id: fact.id.clone(),
                    fact_content: fact.fact.clone(),
                    confidence_before,
                    confidence_after: fact.confidence,
                    new_support_count: fact.support_count,
                    timestamp: chrono::Utc::now(),
                },
            );

            Ok(true)
        } else {
            Ok(false)
        }
    }

    /// Delete a fact (soft delete or hard delete)
    pub fn delete_fact(&self, user_id: &str, fact_id: &str) -> Result<bool> {
        self.fact_store.delete(user_id, fact_id)
    }

    /// Get the fact store for direct access
    pub fn fact_store(&self) -> &Arc<facts::SemanticFactStore> {
        &self.fact_store
    }

    // =========================================================================
    // SHO-118: DECISION LINEAGE GRAPH METHODS
    // =========================================================================

    /// Get the lineage graph for direct access
    pub fn lineage_graph(&self) -> &Arc<lineage::LineageGraph> {
        &self.lineage_graph
    }

    /// Infer and store lineage between a new memory and existing memories
    ///
    /// Called after storing a new memory to automatically detect causal relationships.
    /// Uses entity overlap, temporal proximity, and memory type patterns.
    pub fn infer_lineage_for_memory(
        &self,
        user_id: &str,
        new_memory: &Memory,
        candidate_memories: &[Memory],
    ) -> Result<Vec<LineageEdge>> {
        let mut inferred_edges = Vec::new();

        for candidate in candidate_memories {
            // Backward pass: candidate → new_memory (what caused this memory?)
            if let Some((relation, confidence)) =
                self.lineage_graph.infer_relation(candidate, new_memory)
            {
                if !self
                    .lineage_graph
                    .edge_exists(user_id, &candidate.id, &new_memory.id)?
                {
                    let edge = LineageEdge::inferred(
                        candidate.id.clone(),
                        new_memory.id.clone(),
                        relation,
                        confidence,
                    );
                    self.lineage_graph.store_edge(user_id, &edge)?;
                    inferred_edges.push(edge);
                }
            }

            // Forward pass: new_memory → candidate (what did this memory cause?)
            // This catches retroactive causality: when an earlier memory is stored
            // after a later one (out-of-order ingestion), or when a new memory's
            // type makes it a cause of existing memories (e.g., Learning stored
            // before the Decision it informed).
            if let Some((relation, confidence)) =
                self.lineage_graph.infer_relation(new_memory, candidate)
            {
                if !self
                    .lineage_graph
                    .edge_exists(user_id, &new_memory.id, &candidate.id)?
                {
                    let edge = LineageEdge::inferred(
                        new_memory.id.clone(),
                        candidate.id.clone(),
                        relation,
                        confidence,
                    );
                    self.lineage_graph.store_edge(user_id, &edge)?;
                    inferred_edges.push(edge);
                }
            }
        }

        // Check for branch signal in memory content
        if lineage::LineageGraph::detect_branch_signal(&new_memory.experience.content) {
            self.lineage_graph.ensure_main_branch(user_id)?;
            // Auto-create branch for the pivot
            let branch_name = format!("pivot-{}", chrono::Utc::now().format("%Y%m%d-%H%M%S"));
            match self.lineage_graph.create_branch(
                user_id,
                &branch_name,
                "main",
                new_memory.id.clone(),
                Some(&format!(
                    "Auto-detected pivot: {}",
                    &new_memory
                        .experience
                        .content
                        .chars()
                        .take(80)
                        .collect::<String>()
                )),
            ) {
                Ok(branch) => {
                    tracing::info!(
                        user_id = %user_id,
                        branch = %branch.name,
                        memory_id = %new_memory.id.0,
                        "Auto-created lineage branch from pivot signal"
                    );
                }
                Err(e) => {
                    tracing::debug!("Auto-branch creation failed (non-fatal): {}", e);
                }
            }
        }

        Ok(inferred_edges)
    }

    /// Trace lineage from a memory
    pub fn trace_lineage(
        &self,
        user_id: &str,
        memory_id: &MemoryId,
        direction: TraceDirection,
        max_depth: usize,
    ) -> Result<LineageTrace> {
        self.lineage_graph
            .trace(user_id, memory_id, direction, max_depth)
    }

    /// Find the root cause of a memory
    pub fn find_root_cause(&self, user_id: &str, memory_id: &MemoryId) -> Result<Option<MemoryId>> {
        self.lineage_graph.find_root_cause(user_id, memory_id)
    }

    /// Get lineage statistics
    pub fn lineage_stats(&self, user_id: &str) -> Result<LineageStats> {
        self.lineage_graph.stats(user_id)
    }

    /// Decay facts for all users during maintenance
    ///
    /// Facts decay based on lack of reinforcement. The decay rate is modulated by support_count:
    /// - Higher support = slower decay (fact is well-established)
    /// - Lower support = faster decay (fact is tentative)
    ///
    /// Returns (facts_decayed, facts_deleted)
    fn decay_facts_for_all_users(&self) -> Result<(usize, usize)> {
        use crate::constants::{
            FACT_DECAY_GRACE_DAYS, FACT_DECAY_HALF_LIFE_BASE_DAYS,
            FACT_DECAY_HALF_LIFE_PER_SUPPORT_DAYS,
        };
        const DELETE_CONFIDENCE: f32 = 0.1;

        let now = chrono::Utc::now();
        let mut total_decayed = 0;
        let mut total_deleted = 0;

        let user_ids = self.fact_store.list_users(100)?;

        for user_id in &user_ids {
            let facts = self.fact_store.list(user_id, 10000)?;

            for mut fact in facts {
                let days_since_reinforcement = (now - fact.last_reinforced).num_days();

                // Grace period: no decay at all
                if days_since_reinforcement <= FACT_DECAY_GRACE_DAYS {
                    continue;
                }

                let confidence_before = fact.confidence;

                // Exponential half-life decay: confidence × 0.5^(elapsed / half_life)
                // Half-life grows linearly with support_count — each corroborating source
                // is genuine evidence that the fact is stable knowledge.
                let elapsed = (days_since_reinforcement - FACT_DECAY_GRACE_DAYS) as f64;
                let half_life = FACT_DECAY_HALF_LIFE_BASE_DAYS
                    + (fact.support_count as f64 * FACT_DECAY_HALF_LIFE_PER_SUPPORT_DAYS);
                let decay_factor = (0.5_f64).powf(elapsed / half_life) as f32;
                fact.confidence = (confidence_before * decay_factor).max(0.0);

                // Delete if below threshold
                if fact.confidence < DELETE_CONFIDENCE {
                    // Record deletion event (persists significant events)
                    self.record_consolidation_event_for_user(
                        user_id,
                        ConsolidationEvent::FactDeleted {
                            fact_id: fact.id.clone(),
                            fact_content: fact.fact.clone(),
                            final_confidence: fact.confidence,
                            support_count: fact.support_count,
                            reason: format!("confidence_below_{}", DELETE_CONFIDENCE),
                            timestamp: now,
                        },
                    );

                    self.fact_store.delete(user_id, &fact.id)?;
                    total_deleted += 1;
                } else if (confidence_before - fact.confidence) > 0.001 {
                    // Record decay event (not significant - routine maintenance)
                    self.record_consolidation_event(ConsolidationEvent::FactDecayed {
                        fact_id: fact.id.clone(),
                        fact_content: fact.fact.clone(),
                        confidence_before,
                        confidence_after: fact.confidence,
                        days_since_reinforcement,
                        timestamp: now,
                    });

                    self.fact_store.update(user_id, &fact)?;
                    total_decayed += 1;
                }
            }
        }

        if total_decayed > 0 || total_deleted > 0 {
            tracing::info!(
                facts_decayed = total_decayed,
                facts_deleted = total_deleted,
                "Fact maintenance complete"
            );
        }

        Ok((total_decayed, total_deleted))
    }
}

/// Automatic persistence on drop - ensures vector index and ID mappings survive restarts
///
/// This is CRITICAL for local memory: when the system shuts down (gracefully or via drop),
/// all in-memory state (vector index, ID mappings) must be persisted to disk.
impl Drop for MemorySystem {
    fn drop(&mut self) {
        // Vector index saved via explicit shutdown (save_all_vector_indices)
        // Do NOT save here - Drop fires for temporary instances, overwriting valid saves

        // Flush RocksDB WAL to ensure all writes are durable
        if let Err(e) = self.long_term_memory.flush() {
            tracing::error!("Failed to flush storage on shutdown: {}", e);
        }
    }
}