sqlite_graphrag/commands/

remember.rs

//! Handler for the `remember` CLI subcommand.

use crate::chunking;
use crate::cli::MemoryType;
use crate::entity_type::EntityType;
use crate::errors::AppError;
use crate::i18n::errors_msg;
use crate::output::{self, JsonOutputFormat, RememberResponse};
use crate::paths::AppPaths;
use crate::storage::chunks as storage_chunks;
use crate::storage::connection::{ensure_schema, open_rw};
use crate::storage::entities::{NewEntity, NewRelationship};
use crate::storage::memories::NewMemory;
use crate::storage::{entities, memories, urls as storage_urls, versions};
use serde::Deserialize;

#[derive(clap::Args)]
#[command(after_long_help = "EXAMPLES:\n  \
    # Create a memory with inline body\n  \
    sqlite-graphrag remember --name design-auth --type decision \\\n    \
    --description \"auth design\" --body \"JWT for stateless auth\"\n\n  \
    # Create with curated graph via --graph-stdin\n  \
    echo '{\"body\":\"...\",\"entities\":[],\"relationships\":[]}' | \\\n    \
    sqlite-graphrag remember --name my-mem --type note --description \"desc\" --graph-stdin\n\n  \
    # Enable automatic URL extraction with --graph-stdin (URL-regex only since v1.0.79)\n  \
    echo '{\"body\":\"See https://docs.rs ...\",\"entities\":[],\"relationships\":[]}' | \\\n    \
    sqlite-graphrag remember --name url-test --type note --description \"test\" \\\n    \
    --graph-stdin --enable-ner\n\n  \
    # Idempotent upsert with --force-merge\n  \
    sqlite-graphrag remember --name my-mem --type note --description \"updated\" \\\n    \
    --body \"new content\" --force-merge\n\n\
NOTE:\n  \
    remember does NOT accept positional arguments.\n  \
    Use --body \"text\" for inline content\n  \
    Use --body-file path for file content\n  \
    Use --body-stdin for piped content\n  \
    Use --graph-stdin for JSON with entities and relationships\n\n\
ENTITY TYPES (for --graph-stdin entities, NOT memory --type):\n  \
    concept, tool, person, file, project, decision, incident,\n  \
    organization, location, date, dashboard, issue_tracker, memory\n  \
    WARNING: reference, skill, document, note, user, feedback are\n  \
    MEMORY types only — NOT valid for entities.\n  \
    Mapping: reference→concept, document→file, user→person")]
pub struct RememberArgs {
    /// Memory name in kebab-case (lowercase letters, digits, hyphens).
    /// Acts as unique key within the namespace; collisions trigger merge or rejection.
    #[arg(long)]
    pub name: String,
    #[arg(
        long,
        value_enum,
        long_help = "Memory kind stored in `memories.type`. Required when creating a new memory. Optional with --force-merge: if omitted the existing memory type is inherited. This is NOT the graph `entity_type` used in `--entities-file`. Valid values: user, feedback, project, reference, decision, incident, skill, document, note."
    )]
    pub r#type: Option<MemoryType>,
    /// Short description (≤500 chars) summarizing the memory for use in `list` and `recall` snippets.
    /// Required when creating a new memory. Optional with --force-merge: if omitted the existing description is inherited.
    ///
    /// GAP-SG-33: `allow_hyphen_values` lets a description that begins with a
    /// hyphen (e.g. `"- bullet"`) be accepted as a value instead of being
    /// mistaken for a flag.
    #[arg(long, allow_hyphen_values = true)]
    pub description: Option<String>,
    /// Inline body content. Mutually exclusive with --body-file, --body-stdin, --graph-stdin.
    /// Maximum 512000 bytes; rejected if empty without an external graph.
    ///
    /// GAP-SG-33: `allow_hyphen_values` lets a body that begins with a hyphen
    /// (e.g. a markdown bullet list) be accepted as a value.
    #[arg(
        long,
        allow_hyphen_values = true,
        help = "Inline body content (max 500 KB / 512000 bytes; for larger inputs split into multiple memories or use --body-file)",
        conflicts_with_all = ["body_file", "body_stdin", "graph_stdin"]
    )]
    pub body: Option<String>,
    #[arg(
        long,
        help = "Read body from a file instead of --body",
        conflicts_with_all = ["body", "body_stdin", "graph_stdin"]
    )]
    pub body_file: Option<std::path::PathBuf>,
    /// Read body from stdin until EOF. Useful in pipes (echo "..." | sqlite-graphrag remember ...).
    /// Mutually exclusive with --body, --body-file, --graph-stdin.
    #[arg(
        long,
        conflicts_with_all = ["body", "body_file", "graph_stdin"]
    )]
    pub body_stdin: bool,
    #[arg(
        long,
        help = "JSON file containing entities to associate with this memory"
    )]
    pub entities_file: Option<std::path::PathBuf>,
    #[arg(
        long,
        help = "JSON file containing relationships to associate with this memory"
    )]
    pub relationships_file: Option<std::path::PathBuf>,
    #[arg(
        long,
        help = "Read graph JSON (body + entities + relationships) from stdin",
        conflicts_with_all = [
            "body",
            "body_file",
            "body_stdin",
            "entities_file",
            "relationships_file",
            "graph_file"
        ]
    )]
    pub graph_stdin: bool,
    /// GAP-SG-30: read graph JSON (`{body, entities, relationships}`) from a
    /// FILE instead of stdin, so a curated graph can combine with a body
    /// supplied via --body / --body-file / --body-stdin (which previously
    /// conflicted with --graph-stdin over the single stdin). The file's `body`
    /// field is used only when no other body source is given; otherwise the
    /// body source wins and only the file's entities/relationships are applied.
    #[arg(
        long,
        value_name = "PATH",
        help = "Read graph JSON (body + entities + relationships) from a file (combines with --body/--body-file/--body-stdin)",
        conflicts_with_all = ["graph_stdin", "entities_file", "relationships_file"]
    )]
    pub graph_file: Option<std::path::PathBuf>,
    #[arg(
        long,
        help = "Namespace (env: SQLITE_GRAPHRAG_NAMESPACE, default: global)"
    )]
    pub namespace: Option<String>,
    /// Inline JSON object with arbitrary metadata key-value pairs. Mutually exclusive with --metadata-file.
    #[arg(long)]
    pub metadata: Option<String>,
    #[arg(long, help = "JSON file containing metadata key-value pairs")]
    pub metadata_file: Option<std::path::PathBuf>,
    #[arg(long)]
    pub force_merge: bool,
    #[arg(
        long,
        value_name = "EPOCH_OR_RFC3339",
        value_parser = crate::parsers::parse_expected_updated_at,
        long_help = "Optimistic lock: reject if updated_at does not match. \
Accepts Unix epoch (e.g. 1700000000) or RFC 3339 (e.g. 2026-04-19T12:00:00Z)."
    )]
    pub expected_updated_at: Option<i64>,
    #[arg(
        long,
        env = "SQLITE_GRAPHRAG_ENABLE_NER",
        value_parser = crate::parsers::parse_bool_flexible,
        action = clap::ArgAction::Set,
        num_args = 0..=1,
        default_missing_value = "true",
        default_value = "false",
        help = "Enable automatic URL-regex extraction from body (the GLiNER NER pipeline was removed in v1.0.79)"
    )]
    pub enable_ner: bool,
    #[arg(
        long,
        env = "SQLITE_GRAPHRAG_GLINER_VARIANT",
        default_value = "fp32",
        help = "DEPRECATED: no effect since v1.0.79 (the GLiNER pipeline was removed); accepted for compatibility only"
    )]
    pub gliner_variant: String,
    #[arg(long, hide = true)]
    pub skip_extraction: bool,
    /// Explicitly clear the body content (set to empty string). Required to distinguish
    /// intentional body clearing from accidental omission during --force-merge.
    /// Without this flag, an empty body passed to --force-merge preserves the existing body.
    #[arg(
        long,
        default_value_t = false,
        help = "Explicitly clear body content during --force-merge (without this flag, an empty body is ignored and the existing body is kept)"
    )]
    pub clear_body: bool,
    /// Validate input and report planned actions without persisting.
    #[arg(
        long,
        default_value_t = false,
        help = "Validate input and report planned actions without persisting"
    )]
    pub dry_run: bool,
    /// GAP-SG-37: reject (instead of silently normalizing) when the supplied
    /// --name is not already canonical kebab-case. Use this when the literal
    /// name matters and a silent transform would surprise downstream lookups.
    #[arg(
        long,
        default_value_t = false,
        help = "Reject the write if --name would be normalized to kebab-case (preserve-name guard)"
    )]
    pub strict_name: bool,
    /// GAP-SG-51: with --force-merge, REPLACE the memory's entity/relationship
    /// bindings with the supplied set instead of merging additively. Combined
    /// with an empty `entities`/`relationships` payload this clears all bindings
    /// without deleting the memory.
    #[arg(
        long,
        default_value_t = false,
        help = "With --force-merge, replace (not merge) the memory's graph bindings; empty entities clears them"
    )]
    pub replace_graph: bool,
    /// Optional opaque session identifier for tracing memory provenance across multi-agent runs.
    #[arg(long)]
    pub session_id: Option<String>,
    #[arg(long, value_enum, default_value_t = JsonOutputFormat::Json)]
    pub format: JsonOutputFormat,
    #[arg(long, hide = true, help = "No-op; JSON is always emitted on stdout")]
    pub json: bool,
    #[arg(long, env = "SQLITE_GRAPHRAG_DB_PATH")]
    pub db: Option<String>,
    /// Maximum process RSS in MiB; abort if exceeded during embedding.
    #[arg(long, default_value_t = crate::constants::DEFAULT_MAX_RSS_MB,
          help = "Maximum process RSS in MiB; abort if exceeded during embedding (default: 8192)")]
    pub max_rss_mb: u64,
    /// G42/S3 (v1.0.79): maximum simultaneous LLM embedding subprocesses.
    /// The effective value is further bounded by CPU count and available
    /// RAM (permits = min(N, cpus, ram_livre*0.5/350MB), clamp [1, 32]).
    #[arg(long, default_value_t = 4, value_name = "N",
          value_parser = clap::value_parser!(u64).range(1..=32),
          help = "Maximum simultaneous LLM embedding subprocesses (default: 4, clamp [1,32])")]
    pub llm_parallelism: u64,
}

#[derive(Deserialize, Default)]
#[serde(deny_unknown_fields)]
struct GraphInput {
    #[serde(default)]
    body: Option<String>,
    #[serde(default)]
    entities: Vec<NewEntity>,
    #[serde(default)]
    relationships: Vec<NewRelationship>,
}

fn normalize_and_validate_graph_input(graph: &mut GraphInput) -> Result<(), AppError> {
    for rel in &mut graph.relationships {
        rel.relation = crate::parsers::normalize_relation(&rel.relation);
        if let Err(e) = crate::parsers::validate_relation_format(&rel.relation) {
            return Err(AppError::Validation(format!(
                "{e} for relationship '{}' -> '{}'",
                rel.source, rel.target
            )));
        }
        crate::parsers::warn_if_non_canonical(&rel.relation);
        if !(0.0..=1.0).contains(&rel.strength) {
            return Err(AppError::Validation(format!(
                "invalid strength {} for relationship '{}' -> '{}'; expected value in [0.0, 1.0]",
                rel.strength, rel.source, rel.target
            )));
        }
    }

    Ok(())
}

#[tracing::instrument(skip_all, level = "debug", name = "remember")]
pub fn run(
    args: RememberArgs,
    llm_backend: crate::cli::LlmBackendChoice,
    embedding_backend: crate::cli::EmbeddingBackendChoice,
) -> Result<(), AppError> {
    use crate::constants::*;

    let inicio = std::time::Instant::now();
    let _ = args.format;
    tracing::debug!(target: "remember", name = %args.name, "persisting memory");
    let namespace = crate::namespace::resolve_namespace(args.namespace.as_deref())?;

    // Capture the original `--name` before normalization so the JSON response can
    // surface `name_was_normalized` + `original_name` (B_4 in v1.0.32). Stored as
    // an owned String because `args.name` is moved into the response below.
    let original_name = args.name.clone();

    // Auto-normalize to kebab-case before validation (P2-H).
    // v1.0.20: also trims hyphens at the boundary (including trailing) to avoid rejection
    // after truncation by a long filename ending in a hyphen.
    let normalized_name = {
        let lower = args.name.to_lowercase().replace(['_', ' '], "-");
        let trimmed = lower.trim_matches('-').to_string();
        if trimmed != args.name {
            tracing::warn!(target: "remember",
                original = %args.name,
                normalized = %trimmed,
                "name auto-normalized to kebab-case"
            );
        }
        trimmed
    };
    let name_was_normalized = normalized_name != original_name;

    // GAP-SG-37: when --strict-name is set, refuse to silently rewrite the name.
    // The operator gets the canonical form so they can re-submit it explicitly.
    if args.strict_name && name_was_normalized {
        return Err(AppError::Validation(format!(
            "--strict-name is set but '{original_name}' is not canonical kebab-case; \
             re-run with --name '{normalized_name}' (or drop --strict-name to allow auto-normalization)"
        )));
    }

    if normalized_name.is_empty() {
        return Err(AppError::Validation(
            "name cannot be empty after normalization (input was blank or contained only hyphens/underscores/spaces)".to_string(),
        ));
    }
    if normalized_name.len() > MAX_MEMORY_NAME_LEN {
        return Err(AppError::LimitExceeded(
            crate::i18n::validation::name_length(MAX_MEMORY_NAME_LEN),
        ));
    }

    if normalized_name.starts_with("__") {
        return Err(AppError::Validation(
            crate::i18n::validation::reserved_name(),
        ));
    }

    {
        let slug_re = crate::constants::name_slug_regex();
        if !slug_re.is_match(&normalized_name) {
            return Err(AppError::Validation(crate::i18n::validation::name_kebab(
                &normalized_name,
            )));
        }
    }

    if let Some(ref desc) = args.description {
        if desc.len() > MAX_MEMORY_DESCRIPTION_LEN {
            return Err(AppError::Validation(
                crate::i18n::validation::description_exceeds(MAX_MEMORY_DESCRIPTION_LEN),
            ));
        }
    }

    // GAP-SG-30: capture whether the body comes from an explicit source before
    // `args.body` is moved below; --graph-file only adopts the file's body when
    // no explicit body source was supplied.
    let body_explicitly_provided =
        args.body.is_some() || args.body_file.is_some() || args.body_stdin;

    let mut raw_body = if let Some(b) = args.body {
        b
    } else if let Some(ref path) = args.body_file {
        let file_size = std::fs::metadata(path).map_err(AppError::Io)?.len();
        if file_size > MAX_MEMORY_BODY_LEN as u64 {
            return Err(AppError::BodyTooLarge {
                bytes: file_size,
                limit: MAX_MEMORY_BODY_LEN as u64,
            });
        }
        match std::fs::read_to_string(path) {
            Ok(s) => s,
            Err(e) if e.kind() == std::io::ErrorKind::InvalidData => {
                let bytes = std::fs::read(path).map_err(AppError::Io)?;
                tracing::warn!(target: "remember", "body file contains invalid UTF-8; replacing invalid sequences");
                String::from_utf8_lossy(&bytes).into_owned()
            }
            Err(e) => return Err(AppError::Io(e)),
        }
    } else if args.body_stdin || args.graph_stdin {
        crate::stdin_helper::read_stdin_with_timeout(60)?
    } else {
        String::new()
    };

    let mut entities_provided_externally =
        args.entities_file.is_some() || args.relationships_file.is_some();

    let mut graph = GraphInput::default();
    if let Some(path) = args.entities_file {
        let file_size = std::fs::metadata(&path).map_err(AppError::Io)?.len();
        if file_size > MAX_MEMORY_BODY_LEN as u64 {
            return Err(AppError::BodyTooLarge {
                bytes: file_size,
                limit: MAX_MEMORY_BODY_LEN as u64,
            });
        }
        let content = std::fs::read_to_string(&path).map_err(AppError::Io)?;
        // v1.1.1 (P7): boundary validation with context — an invalid
        // entity_type surfaces the FromStr message (13 valid values + hints)
        // as a Validation error instead of a bare Json error (exit 20).
        graph.entities = serde_json::from_str(&content)
            .map_err(|e| AppError::Validation(format!("invalid JSON in --entities-file: {e}")))?;
    }
    if let Some(path) = args.relationships_file {
        let file_size = std::fs::metadata(&path).map_err(AppError::Io)?.len();
        if file_size > MAX_MEMORY_BODY_LEN as u64 {
            return Err(AppError::BodyTooLarge {
                bytes: file_size,
                limit: MAX_MEMORY_BODY_LEN as u64,
            });
        }
        let content = std::fs::read_to_string(&path).map_err(AppError::Io)?;
        graph.relationships = serde_json::from_str(&content).map_err(|e| {
            AppError::Validation(format!("invalid JSON in --relationships-file: {e}"))
        })?;
    }
    if args.graph_stdin {
        graph = serde_json::from_str::<GraphInput>(&raw_body).map_err(|e| {
            AppError::Validation(format!("invalid JSON payload on --graph-stdin: {e}"))
        })?;
        raw_body = graph.body.take().unwrap_or_default();
    }
    if args.graph_stdin && !graph.entities.is_empty() {
        entities_provided_externally = true;
    }
    // GAP-SG-30: graph from a file, combinable with any body source. Conflicts
    // with --graph-stdin/--entities-file/--relationships-file (enforced by clap),
    // so `graph` is still empty here when --graph-file is set.
    if let Some(path) = args.graph_file {
        let file_size = std::fs::metadata(&path).map_err(AppError::Io)?.len();
        if file_size > MAX_MEMORY_BODY_LEN as u64 {
            return Err(AppError::BodyTooLarge {
                bytes: file_size,
                limit: MAX_MEMORY_BODY_LEN as u64,
            });
        }
        let content = std::fs::read_to_string(&path).map_err(AppError::Io)?;
        let mut gf = serde_json::from_str::<GraphInput>(&content)
            .map_err(|e| AppError::Validation(format!("invalid JSON in --graph-file: {e}")))?;
        graph.entities = gf.entities;
        graph.relationships = gf.relationships;
        if !body_explicitly_provided {
            raw_body = gf.body.take().unwrap_or_default();
        }
        if !graph.entities.is_empty() {
            entities_provided_externally = true;
        }
    }

    if graph.entities.len() > max_entities_per_memory() {
        return Err(AppError::LimitExceeded(errors_msg::entity_limit_exceeded(
            max_entities_per_memory(),
        )));
    }
    let mut relationships_truncated = false;
    let rel_cap = max_relationships_per_memory();
    if graph.relationships.len() > rel_cap {
        tracing::warn!(target: "remember",
            count = graph.relationships.len(),
            cap = rel_cap,
            "truncating relationships to cap"
        );
        graph.relationships.truncate(rel_cap);
        relationships_truncated = true;
    }
    normalize_and_validate_graph_input(&mut graph)?;

    if raw_body.len() > MAX_MEMORY_BODY_LEN {
        return Err(AppError::BodyTooLarge {
            bytes: raw_body.len() as u64,
            limit: MAX_MEMORY_BODY_LEN as u64,
        });
    }

    // v1.0.22 P1: reject empty or whitespace-only body when no external graph is provided.
    // Without this check, empty embeddings would be persisted, breaking recall semantics.
    // GAP-08: skip this guard when --force-merge without --clear-body; the existing body
    // will be preserved from the database, so the effective body will not be empty.
    let body_will_be_preserved = args.force_merge && raw_body.trim().is_empty() && !args.clear_body;
    if !entities_provided_externally
        && graph.entities.is_empty()
        && raw_body.trim().is_empty()
        && !body_will_be_preserved
        && !args.clear_body
    {
        return Err(AppError::Validation(crate::i18n::validation::empty_body()));
    }

    let metadata: serde_json::Value = if let Some(m) = args.metadata {
        serde_json::from_str(&m)?
    } else if let Some(path) = args.metadata_file {
        let file_size = std::fs::metadata(&path).map_err(AppError::Io)?.len();
        if file_size > MAX_MEMORY_BODY_LEN as u64 {
            return Err(AppError::BodyTooLarge {
                bytes: file_size,
                limit: MAX_MEMORY_BODY_LEN as u64,
            });
        }
        let content = std::fs::read_to_string(&path).map_err(AppError::Io)?;
        serde_json::from_str(&content)?
    } else {
        serde_json::json!({})
    };

    let mut body_hash = blake3::hash(raw_body.as_bytes()).to_hex().to_string();
    let mut snippet: String = raw_body.chars().take(200).collect();

    let paths = AppPaths::resolve(args.db.as_deref())?;
    paths.ensure_dirs()?;

    // v1.0.20: use .trim().is_empty() to reject bodies that are only whitespace.
    let mut extraction_method: Option<String> = None;
    let mut extracted_urls: Vec<crate::extraction::ExtractedUrl> = Vec::with_capacity(4);
    if args.enable_ner && args.skip_extraction {
        return Err(AppError::Validation(
            "--enable-ner and --skip-extraction are mutually exclusive; remove one".to_string(),
        ));
    }
    if args.skip_extraction && !args.enable_ner {
        // v1.0.74: revert to v1.0.45 hidden no-op behavior. The v1.0.67
        // commit (9ddb17b) promoted this to a hard validation error, which
        // broke the "kept as a hidden no-op for backwards compatibility"
        // promise documented in CHANGELOG v1.0.45 and started failing
        // 5+ CI jobs whose E2E tests use this flag to skip the
        // GLiNER-ONNX model download in CI environments.
        tracing::warn!(
            "--skip-extraction is deprecated since v1.0.45 and has no effect (NER is disabled by default); remove this flag to silence the warning"
        );
    }
    // v1.0.79: --gliner-variant is a no-op kept for compatibility; a
    // non-default value signals the caller still expects the removed
    // GLiNER pipeline, so warn explicitly.
    if args.gliner_variant != "fp32" {
        tracing::warn!(
            "--gliner-variant is deprecated and has no effect since v1.0.79 (the GLiNER pipeline was removed); --enable-ner performs URL-regex extraction only"
        );
    }
    let gliner_variant: crate::extraction::GlinerVariant = match args.gliner_variant.as_str() {
        "int8" => crate::extraction::GlinerVariant::Int8,
        _ => crate::extraction::GlinerVariant::Fp32,
    };
    if args.enable_ner && graph.entities.is_empty() && !raw_body.trim().is_empty() {
        match crate::extraction::extract_graph_auto(&raw_body, &paths, gliner_variant) {
            Ok(extracted) => {
                // v1.0.76: ExtractionResult is URL + entity + elapsed_ms;
                // the LLM ExtractionBackend returns typed relationships
                // separately. The default build is URL-only extraction.
                extraction_method = Some("url-regex".to_string());
                extracted_urls = extracted.urls;
                // Convert ExtractedEntity → NewEntity (no offsets,
                // type defaults to Concept).
                graph.entities = extracted
                    .entities
                    .into_iter()
                    .map(|e| NewEntity {
                        name: e.name,
                        entity_type: crate::entity_type::EntityType::Concept,
                        description: None,
                    })
                    .collect();
                graph.relationships.clear();
                relationships_truncated = false;

                if graph.entities.len() > max_entities_per_memory() {
                    graph.entities.truncate(max_entities_per_memory());
                }
                if graph.relationships.len() > max_relationships_per_memory() {
                    relationships_truncated = true;
                    graph.relationships.truncate(max_relationships_per_memory());
                }
                normalize_and_validate_graph_input(&mut graph)?;
            }
            Err(e) => {
                tracing::warn!(target: "remember", error = %e, "auto-extraction failed, graceful degradation");
                extraction_method = Some("none:extraction-failed".to_string());
            }
        }
    }

    let mut conn = open_rw(&paths.db)?;
    ensure_schema(&mut conn)?;

    // --dry-run: emit planned action without any DB writes and return.
    if args.dry_run {
        let existing = memories::find_by_name(&conn, &namespace, &normalized_name)?;
        let planned_action = if existing.is_some() && args.force_merge {
            "would_update"
        } else {
            "would_create"
        };
        output::emit_json(&serde_json::json!({
            "dry_run": true,
            "name": normalized_name,
            "namespace": namespace,
            "planned_action": planned_action,
        }))?;
        return Ok(());
    }

    {
        use crate::constants::MAX_NAMESPACES_ACTIVE;
        let active_count: u32 = conn.query_row(
            "SELECT COUNT(DISTINCT namespace) FROM memories WHERE deleted_at IS NULL",
            [],
            |r| r.get::<_, i64>(0).map(|v| v as u32),
        )?;
        let ns_exists: bool = conn.query_row(
            "SELECT EXISTS(SELECT 1 FROM memories WHERE namespace = ?1 AND deleted_at IS NULL)",
            rusqlite::params![namespace],
            |r| r.get::<_, i64>(0).map(|v| v > 0),
        )?;
        if !ns_exists && active_count >= MAX_NAMESPACES_ACTIVE {
            return Err(AppError::NamespaceError(format!(
                "active namespace limit of {MAX_NAMESPACES_ACTIVE} reached while trying to create '{namespace}'"
            )));
        }
    }

    // M7: detect soft-deleted memory before the standard duplicate check.
    if let Some((sd_id, true)) =
        memories::find_by_name_any_state(&conn, &namespace, &normalized_name)?
    {
        if args.force_merge {
            memories::clear_deleted_at(&conn, sd_id)?;
        } else {
            return Err(AppError::Duplicate(
                errors_msg::duplicate_memory_soft_deleted(&normalized_name, &namespace),
            ));
        }
    }

    let existing_memory = memories::find_by_name(&conn, &namespace, &normalized_name)?;
    if existing_memory.is_some() && !args.force_merge {
        return Err(AppError::Duplicate(errors_msg::duplicate_memory(
            &normalized_name,
            &namespace,
        )));
    }

    // GAP-10: resolve type and description.
    // For CREATE path (new memory): both are required.
    // For UPDATE path (--force-merge on existing memory): inherit from existing row when omitted.
    let (resolved_type, resolved_description) = if existing_memory.is_none() {
        // CREATE path — both fields are mandatory.
        let t = args.r#type.ok_or_else(|| {
            AppError::Validation(
                "--type and --description are required when creating a new memory".to_string(),
            )
        })?;
        let d = args.description.clone().ok_or_else(|| {
            AppError::Validation(
                "--type and --description are required when creating a new memory".to_string(),
            )
        })?;
        (t.as_str().to_string(), d)
    } else {
        // UPDATE path (--force-merge) — inherit missing fields from stored row.
        let existing_row = memories::read_by_name(&conn, &namespace, &normalized_name)?
            .ok_or_else(|| {
                AppError::NotFound(format!(
                    "memory '{normalized_name}' not found in namespace '{namespace}'"
                ))
            })?;
        let t = args
            .r#type
            .map(|v| v.as_str().to_string())
            .unwrap_or_else(|| existing_row.memory_type.clone());
        let d = args
            .description
            .clone()
            .unwrap_or_else(|| existing_row.description.clone());
        (t, d)
    };

    // GAP-08/GAP-09: protect existing body from accidental destruction during --force-merge.
    // When the caller omits a body (or passes an empty one) without --clear-body, silently
    // preserve the existing body from the database.  This prevents a common scripting mistake
    // where a cron job updates metadata fields and inadvertently wipes the stored content.
    if body_will_be_preserved {
        if let Some(existing_row) = memories::read_by_name(&conn, &namespace, &normalized_name)? {
            if !existing_row.body.is_empty() {
                tracing::debug!(target: "remember",
                    name = %normalized_name,
                    "GAP-08: empty body with --force-merge and no --clear-body; preserving existing body"
                );
                raw_body = existing_row.body;
                body_hash = blake3::hash(raw_body.as_bytes()).to_hex().to_string();
                snippet = raw_body.chars().take(200).collect();
            }
        }
    }

    let duplicate_hash_id = memories::find_by_hash(&conn, &namespace, &body_hash)?;

    output::emit_progress_i18n(
        &format!(
            "Remember stage: validated input; available memory {} MB",
            crate::memory_guard::available_memory_mb()
        ),
        &format!(
            "Stage remember: input validated; available memory {} MB",
            crate::memory_guard::available_memory_mb()
        ),
    );

    let model_max_length = crate::tokenizer::get_model_max_length();
    let total_passage_tokens = crate::tokenizer::count_passage_tokens(&raw_body)?;
    let chunks_info = chunking::split_into_chunks_hierarchical(&raw_body);
    let chunks_created = chunks_info.len();
    // GAP-SG-40: `chunks_persisted` is no longer a pre-commit estimate. It is
    // read back from `memory_chunks` AFTER the transaction commits (see below)
    // so the reported count matches the observable database state. Single-chunk
    // bodies store inline in the memories row and append no chunk rows.

    output::emit_progress_i18n(
        &format!(
            "Remember stage: tokenizer counted {total_passage_tokens} passage tokens (model max {model_max_length}); chunking produced {} chunks; process RSS {} MB",
            chunks_created,
            crate::memory_guard::current_process_memory_mb().unwrap_or(0)
        ),
        &format!(
            "Stage remember: tokenizer counted {total_passage_tokens} passage tokens (model max {model_max_length}); chunking produced {} chunks; process RSS {} MB",
            chunks_created,
            crate::memory_guard::current_process_memory_mb().unwrap_or(0)
        ),
    );

    if chunks_created > crate::constants::REMEMBER_MAX_SAFE_MULTI_CHUNKS {
        return Err(AppError::TooManyChunks {
            chunks: chunks_created,
            limit: crate::constants::REMEMBER_MAX_SAFE_MULTI_CHUNKS,
        });
    }

    output::emit_progress_i18n("Computing embedding...", "Calculando embedding...");
    let mut chunk_embeddings_cache: Option<Vec<Vec<f32>>> = None;

    // v1.0.84 (ADR-0042): extrai o backend que efetivamente executou o
    // embedding da passagem (ou do batch em chunks) para popular
    // `backend_invoked` no envelope de resposta.
    let skip_embed = crate::embedder::should_skip_embedding_on_failure();
    let (embedding, backend_invoked_passage): (Option<Vec<f32>>, Option<&str>) = if chunks_info
        .len()
        == 1
    {
        match crate::embedder::embed_passage_with_embedding_choice(
            &paths.models,
            &raw_body,
            embedding_backend,
            llm_backend,
        ) {
            Ok((v, k)) => (Some(v), Some(k.as_str())),
            Err(AppError::Validation(msg)) => return Err(AppError::Validation(msg)),
            Err(e) if skip_embed => {
                tracing::warn!(error = %e, "embedding failed; --skip-embedding-on-failure active, persisting without embedding");
                (None, None)
            }
            Err(e) => return Err(e),
        }
    } else {
        let chunk_texts: Vec<String> = chunks_info
            .iter()
            .map(|c| chunking::chunk_text(&raw_body, c).to_string())
            .collect();
        // G42/S2+S3 (v1.0.79): chunks are embedded in dim-adaptive
        // batches per LLM call (G44: clamp(base*64/dim, 1, base)), with up to
        // --llm-parallelism bounded subprocesses in flight. The old
        // serial loop spent SUM(items) wall time; the fan-out spends
        // roughly MAX(batch).
        output::emit_progress_i18n(
            &format!(
                "Embedding {} chunks in parallel batches (parallelism {})...",
                chunks_info.len(),
                args.llm_parallelism
            ),
            &format!(
                "Embedding {} chunks em lotes paralelos (paralelismo {})...",
                chunks_info.len(),
                args.llm_parallelism
            ),
        );
        if let Some(rss) = crate::memory_guard::current_process_memory_mb() {
            if rss > args.max_rss_mb {
                tracing::error!(target: "remember",
                    rss_mb = rss,
                    max_rss_mb = args.max_rss_mb,
                    "RSS exceeded --max-rss-mb threshold; aborting to prevent system instability"
                );
                return Err(AppError::LowMemory {
                    available_mb: crate::memory_guard::available_memory_mb(),
                    required_mb: args.max_rss_mb,
                });
            }
        }
        match crate::embedder::embed_passages_parallel_with_embedding_choice(
            &paths.models,
            &chunk_texts,
            args.llm_parallelism as usize,
            crate::embedder::chunk_embed_batch_size(),
            embedding_backend,
            llm_backend,
        ) {
            Ok(chunk_embeddings) => {
                output::emit_progress_i18n(
                    &format!(
                        "Remember stage: chunk embeddings complete; process RSS {} MB",
                        crate::memory_guard::current_process_memory_mb().unwrap_or(0)
                    ),
                    &format!(
                        "Stage remember: chunk embeddings completed; process RSS {} MB",
                        crate::memory_guard::current_process_memory_mb().unwrap_or(0)
                    ),
                );
                let aggregated = chunking::aggregate_embeddings(&chunk_embeddings);
                chunk_embeddings_cache = Some(chunk_embeddings);
                (Some(aggregated), None)
            }
            Err(e) if skip_embed => {
                tracing::warn!(error = %e, "chunk embedding failed; --skip-embedding-on-failure active, persisting without embedding");
                (None, None)
            }
            Err(e) => return Err(e),
        }
    };
    let body_for_storage = raw_body;

    let memory_type = resolved_type.as_str();
    let new_memory = NewMemory {
        namespace: namespace.clone(),
        name: normalized_name.clone(),
        memory_type: memory_type.to_string(),
        description: resolved_description.clone(),
        body: body_for_storage,
        body_hash: body_hash.clone(),
        session_id: args.session_id.clone(),
        source: "agent".to_string(),
        metadata,
    };

    let mut warnings = Vec::with_capacity(4);
    let mut entities_persisted = 0usize;
    let mut relationships_persisted = 0usize;

    // G42/S2+A4 (v1.0.79): entity names are SHORT texts — they get their
    // own batch profile (25 per LLM call) instead of one subprocess per
    // 3-15 byte name (21 names used to cost ~12 minutes, 46% of the
    // measured remember total).
    let entity_texts: Vec<String> = graph
        .entities
        .iter()
        .map(|entity| match &entity.description {
            Some(desc) => format!("{} {}", entity.name, desc),
            None => entity.name.clone(),
        })
        .collect();
    // G56 (v1.0.80): route entity-name embedding through the in-process
    // cache. Repeated `remember` invocations within one CLI process — and
    // re-embedded entities inside a single batch — skip the LLM call
    // entirely when the (model, text) pair was already produced. The
    // chunk body embedding below still uses `embed_passages_parallel_local`
    // because chunks are unique per memory and the cache hit rate is
    // effectively zero.
    let (graph_entity_embeddings, embed_cache_stats) =
        match crate::embedder::embed_entity_texts_cached(
            &paths.models,
            &entity_texts,
            args.llm_parallelism as usize,
            embedding_backend,
            llm_backend,
        ) {
            Ok(r) => r,
            Err(e) if skip_embed => {
                tracing::warn!(error = %e, "entity embedding failed; --skip-embedding-on-failure active");
                let empty: Vec<Vec<f32>> = entity_texts.iter().map(|_| vec![]).collect();
                (empty, crate::embedder::EmbedCacheStats::default())
            }
            Err(e) => return Err(e),
        };
    if embed_cache_stats.hits > 0 {
        tracing::debug!(
            hits = embed_cache_stats.hits,
            misses = embed_cache_stats.misses,
            requested = embed_cache_stats.requested,
            "G56: entity embed cache hit (remember)"
        );
    }

    let tx = conn.transaction_with_behavior(rusqlite::TransactionBehavior::Immediate)?;

    let mut skip_reindex = false;
    let (memory_id, action, version) = match existing_memory {
        Some((existing_id, _updated_at, _current_version)) => {
            if let Some(hash_id) = duplicate_hash_id {
                if hash_id != existing_id {
                    warnings.push(format!(
                        "identical body already exists as memory id {hash_id}"
                    ));
                }
            }

            // C1 fix: capture old values for FTS5 sync before update
            let (old_fts_name, old_fts_desc, old_fts_body): (String, String, String) = tx
                .query_row(
                    "SELECT name, description, body FROM memories WHERE id = ?1",
                    rusqlite::params![existing_id],
                    |r| Ok((r.get(0)?, r.get(1)?, r.get(2)?)),
                )?;

            // G15: skip re-indexing when body hash matches (common in --force-merge loops)
            let existing_body_hash: Option<String> = tx
                .query_row(
                    "SELECT body_hash FROM memories WHERE id = ?1",
                    rusqlite::params![existing_id],
                    |r| r.get(0),
                )
                .ok();
            let body_unchanged = existing_body_hash.as_deref() == Some(&body_hash);
            skip_reindex = body_unchanged;
            if !body_unchanged {
                storage_chunks::delete_chunks(&tx, existing_id)?;
            }

            let next_v = versions::next_version(&tx, existing_id)?;
            memories::update(&tx, existing_id, &new_memory, args.expected_updated_at)?;

            // C1 fix: sync FTS5 external-content index after update
            // (trg_fts_au trigger is absent by design due to sqlite-vec conflict)
            memories::sync_fts_after_update(
                &tx,
                existing_id,
                &old_fts_name,
                &old_fts_desc,
                &old_fts_body,
                &normalized_name,
                &resolved_description,
                &new_memory.body,
            )?;

            versions::insert_version(
                &tx,
                existing_id,
                next_v,
                &normalized_name,
                memory_type,
                &resolved_description,
                &new_memory.body,
                &serde_json::to_string(&new_memory.metadata)?,
                None,
                "edit",
            )?;
            if !body_unchanged {
                if let Some(ref emb) = embedding {
                    memories::upsert_vec(
                        &tx,
                        existing_id,
                        &namespace,
                        memory_type,
                        emb,
                        &normalized_name,
                        &snippet,
                    )?;
                }
            }
            (existing_id, "updated".to_string(), next_v)
        }
        None => {
            if let Some(hash_id) = duplicate_hash_id {
                warnings.push(format!(
                    "identical body already exists as memory id {hash_id}"
                ));
            }
            let id = memories::insert(&tx, &new_memory)?;
            versions::insert_version(
                &tx,
                id,
                1,
                &normalized_name,
                memory_type,
                &resolved_description,
                &new_memory.body,
                &serde_json::to_string(&new_memory.metadata)?,
                None,
                "create",
            )?;
            if let Some(ref emb) = embedding {
                memories::upsert_vec(
                    &tx,
                    id,
                    &namespace,
                    memory_type,
                    emb,
                    &normalized_name,
                    &snippet,
                )?;
            }
            (id, "created".to_string(), 1)
        }
    };

    // GAP-SG-51: when --force-merge --replace-graph updates an existing memory,
    // clear its prior entity/relationship bindings BEFORE re-linking the supplied
    // set. With an empty `entities`/`relationships` payload this zeroes the graph
    // for that memory without a `forget`. New bindings (if any) are linked by the
    // block further below.
    if args.replace_graph && action == "updated" {
        let (e_removed, r_removed) = entities::clear_memory_graph_bindings(&tx, memory_id)?;
        if e_removed + r_removed > 0 {
            warnings.push(format!(
                "--replace-graph cleared {e_removed} entity binding(s) and {r_removed} relationship binding(s) before re-linking"
            ));
        }
    }

    if chunks_info.len() > 1 && !skip_reindex {
        storage_chunks::insert_chunk_slices(&tx, memory_id, &new_memory.body, &chunks_info)?;

        if let Some(chunk_embeddings) = chunk_embeddings_cache.take() {
            for (i, emb) in chunk_embeddings.iter().enumerate() {
                storage_chunks::upsert_chunk_vec(&tx, i as i64, memory_id, i as i32, emb)?;
            }
        }
        output::emit_progress_i18n(
            &format!(
                "Remember stage: persisted chunk vectors; process RSS {} MB",
                crate::memory_guard::current_process_memory_mb().unwrap_or(0)
            ),
            &format!(
                "Etapa remember: vetores de chunks persistidos; RSS do processo {} MB",
                crate::memory_guard::current_process_memory_mb().unwrap_or(0)
            ),
        );
    }

    if !graph.entities.is_empty() || !graph.relationships.is_empty() {
        for entity in &graph.entities {
            let entity_id = entities::upsert_entity(&tx, &namespace, entity)?;
            let entity_embedding = &graph_entity_embeddings[entities_persisted];
            entities::upsert_entity_vec(
                &tx,
                entity_id,
                &namespace,
                entity.entity_type,
                entity_embedding,
                &entity.name,
            )?;
            entities::link_memory_entity(&tx, memory_id, entity_id)?;
            entities_persisted += 1;
        }
        let entity_types: std::collections::HashMap<&str, EntityType> = graph
            .entities
            .iter()
            .map(|entity| (entity.name.as_str(), entity.entity_type))
            .collect();

        let mut affected_entity_ids: std::collections::HashSet<i64> =
            std::collections::HashSet::new();
        for entity in &graph.entities {
            if let Some(eid) = entities::find_entity_id(&tx, &namespace, &entity.name)? {
                affected_entity_ids.insert(eid);
            }
        }

        for rel in &graph.relationships {
            let source_entity = NewEntity {
                name: rel.source.clone(),
                entity_type: entity_types
                    .get(rel.source.as_str())
                    .copied()
                    .unwrap_or(EntityType::Concept),
                description: None,
            };
            let target_entity = NewEntity {
                name: rel.target.clone(),
                entity_type: entity_types
                    .get(rel.target.as_str())
                    .copied()
                    .unwrap_or(EntityType::Concept),
                description: None,
            };
            let source_id = entities::upsert_entity(&tx, &namespace, &source_entity)?;
            let target_id = entities::upsert_entity(&tx, &namespace, &target_entity)?;
            let rel_id = entities::upsert_relationship(&tx, &namespace, source_id, target_id, rel)?;
            entities::link_memory_relationship(&tx, memory_id, rel_id)?;
            affected_entity_ids.insert(source_id);
            affected_entity_ids.insert(target_id);
            relationships_persisted += 1;
        }

        for &eid in &affected_entity_ids {
            entities::recalculate_degree(&tx, eid)?;
        }
    }
    tx.commit()?;

    // GAP-SG-40: read back the real chunk-row count now that the write is
    // durable, so `chunks_persisted` reflects observable state (0 for inline
    // single-chunk bodies, the exact row count for multi-chunk bodies).
    let chunks_persisted = storage_chunks::count_for_memory(&conn, memory_id)?;

    // GAP-SG-44: confirm the memory has a persisted embedding vector. A missing
    // vector (embedding step failed/skipped) makes the memory unsearchable
    // silently; surface it as a warning recommending re-embed instead of leaving
    // `health` to report `vec_memories_missing` later.
    if !new_memory.body.trim().is_empty() {
        let has_vec: bool = conn
            .query_row(
                "SELECT EXISTS(SELECT 1 FROM memory_embeddings WHERE memory_id = ?1)",
                rusqlite::params![memory_id],
                |r| r.get::<_, i64>(0).map(|v| v > 0),
            )
            .unwrap_or(false);
        if !has_vec {
            tracing::warn!(target: "remember",
                memory_id,
                name = %normalized_name,
                "memory persisted without an embedding vector; recall will be degraded until re-embedded"
            );
            warnings.push(
                "memory persisted without an embedding vector; run `enrich --operation re-embed` to make it searchable"
                    .to_string(),
            );
        }
    }

    // GAP-SG-13: when --force-merge UPDATES an existing memory its body/graph may
    // have changed, so drop any stale enrich-queue sidecar entry keyed to it. The
    // next enrich run re-scans it cleanly. Best-effort; no-op when the queue file
    // is absent.
    if action == "updated" {
        crate::commands::enrich::cleanup_queue_entry(&paths.db, memory_id, &normalized_name);
    }

    // v1.0.24 P0-2: persist URLs in a dedicated table, outside the main transaction.
    // Failures do not propagate — non-critical path with graceful degradation.
    let urls_persisted = if !extracted_urls.is_empty() {
        let url_entries: Vec<storage_urls::MemoryUrl> = extracted_urls
            .into_iter()
            .map(|u| storage_urls::MemoryUrl {
                url: u.url,
                offset: Some(u.start as i64),
            })
            .collect();
        storage_urls::insert_urls(&conn, memory_id, &url_entries)
    } else {
        0
    };

    conn.execute_batch("PRAGMA wal_checkpoint(TRUNCATE);")?;

    let created_at_epoch = chrono::Utc::now().timestamp();
    let created_at_iso = crate::tz::format_iso(chrono::Utc::now());

    output::emit_json(&RememberResponse {
        memory_id,
        // Persist the normalized (kebab-case) slug as `name` since that is the
        // storage key. The original input is exposed via `original_name` only
        // when normalization actually changed something (B_4 in v1.0.32).
        name: normalized_name.clone(),
        namespace,
        action: action.clone(),
        operation: action,
        version,
        entities_persisted,
        relationships_persisted,
        relationships_truncated,
        chunks_created,
        chunks_persisted,
        urls_persisted,
        extraction_method,
        merged_into_memory_id: None,
        warnings,
        created_at: created_at_epoch,
        created_at_iso,
        elapsed_ms: inicio.elapsed().as_millis() as u64,
        name_was_normalized,
        original_name: name_was_normalized.then_some(original_name),
        backend_invoked: backend_invoked_passage,
    })?;

    Ok(())
}

#[cfg(test)]
mod tests {
    use crate::output::RememberResponse;

    /// GAP-SG-37: replicates the `--strict-name` guard predicate so the
    /// reject-on-normalization decision is unit-testable without a DB.
    fn strict_name_rejects(strict: bool, name_was_normalized: bool) -> bool {
        strict && name_was_normalized
    }

    #[test]
    fn strict_name_rejects_only_when_name_would_change() {
        assert!(
            strict_name_rejects(true, true),
            "strict + changed must reject"
        );
        assert!(
            !strict_name_rejects(true, false),
            "strict + canonical passes"
        );
        assert!(
            !strict_name_rejects(false, true),
            "non-strict always passes"
        );
        assert!(!strict_name_rejects(false, false));
    }

    // GAP-SG-37/SG-51: --strict-name and --replace-graph must parse on remember.
    #[test]
    fn remember_parses_strict_name_and_replace_graph_flags() {
        use crate::cli::{Cli, Commands};
        use clap::Parser;
        let cli = Cli::try_parse_from([
            "sqlite-graphrag",
            "remember",
            "--name",
            "my-mem",
            "--type",
            "note",
            "--description",
            "d",
            "--body",
            "b",
            "--strict-name",
            "--replace-graph",
            "--force-merge",
        ])
        .expect("parse");
        match cli.command {
            Some(Commands::Remember(a)) => {
                assert!(a.strict_name);
                assert!(a.replace_graph);
                assert!(a.force_merge);
            }
            other => panic!("expected remember, got {other:?}"),
        }
    }

    #[test]
    fn remember_response_serializes_required_fields() {
        let resp = RememberResponse {
            memory_id: 42,
            name: "minha-mem".to_string(),
            namespace: "global".to_string(),
            action: "created".to_string(),
            operation: "created".to_string(),
            version: 1,
            entities_persisted: 0,
            relationships_persisted: 0,
            relationships_truncated: false,
            chunks_created: 1,
            chunks_persisted: 0,
            urls_persisted: 0,
            extraction_method: None,
            merged_into_memory_id: None,
            warnings: vec![],
            created_at: 1_705_320_000,
            created_at_iso: "2024-01-15T12:00:00Z".to_string(),
            elapsed_ms: 55,
            name_was_normalized: false,
            original_name: None,
            backend_invoked: None,
        };

        let json = serde_json::to_value(&resp).expect("serialization failed");
        assert_eq!(json["memory_id"], 42);
        assert_eq!(json["action"], "created");
        assert_eq!(json["operation"], "created");
        assert_eq!(json["version"], 1);
        assert_eq!(json["elapsed_ms"], 55u64);
        assert!(json["warnings"].is_array());
        assert!(json["merged_into_memory_id"].is_null());
    }

    #[test]
    fn remember_response_action_e_operation_sao_aliases() {
        let resp = RememberResponse {
            memory_id: 1,
            name: "mem".to_string(),
            namespace: "global".to_string(),
            action: "updated".to_string(),
            operation: "updated".to_string(),
            version: 2,
            entities_persisted: 3,
            relationships_persisted: 1,
            relationships_truncated: false,
            extraction_method: None,
            chunks_created: 2,
            chunks_persisted: 2,
            urls_persisted: 0,
            merged_into_memory_id: None,
            warnings: vec![],
            created_at: 0,
            created_at_iso: "1970-01-01T00:00:00Z".to_string(),
            elapsed_ms: 0,
            name_was_normalized: false,
            original_name: None,
            backend_invoked: None,
        };

        let json = serde_json::to_value(&resp).expect("serialization failed");
        assert_eq!(
            json["action"], json["operation"],
            "action e operation devem ser iguais"
        );
        assert_eq!(json["entities_persisted"], 3);
        assert_eq!(json["relationships_persisted"], 1);
        assert_eq!(json["chunks_created"], 2);
    }

    #[test]
    fn remember_response_warnings_lista_mensagens() {
        let resp = RememberResponse {
            memory_id: 5,
            name: "dup-mem".to_string(),
            namespace: "global".to_string(),
            action: "created".to_string(),
            operation: "created".to_string(),
            version: 1,
            entities_persisted: 0,
            extraction_method: None,
            relationships_persisted: 0,
            relationships_truncated: false,
            chunks_created: 1,
            chunks_persisted: 0,
            urls_persisted: 0,
            merged_into_memory_id: None,
            warnings: vec!["identical body already exists as memory id 3".to_string()],
            created_at: 0,
            created_at_iso: "1970-01-01T00:00:00Z".to_string(),
            elapsed_ms: 10,
            name_was_normalized: false,
            original_name: None,
            backend_invoked: None,
        };

        let json = serde_json::to_value(&resp).expect("serialization failed");
        let warnings = json["warnings"]
            .as_array()
            .expect("warnings deve ser array");
        assert_eq!(warnings.len(), 1);
        assert!(warnings[0].as_str().unwrap().contains("identical body"));
    }

    #[test]
    fn invalid_name_reserved_prefix_returns_validation_error() {
        use crate::errors::AppError;
        // Validates the rejection logic for names with the "__" prefix directly
        let nome = "__reservado";
        let resultado: Result<(), AppError> = if nome.starts_with("__") {
            Err(AppError::Validation(
                crate::i18n::validation::reserved_name(),
            ))
        } else {
            Ok(())
        };
        assert!(resultado.is_err());
        if let Err(AppError::Validation(msg)) = resultado {
            assert!(!msg.is_empty());
        }
    }

    #[test]
    fn name_too_long_returns_validation_error() {
        use crate::errors::AppError;
        let nome_longo = "a".repeat(crate::constants::MAX_MEMORY_NAME_LEN + 1);
        let resultado: Result<(), AppError> =
            if nome_longo.is_empty() || nome_longo.len() > crate::constants::MAX_MEMORY_NAME_LEN {
                Err(AppError::Validation(crate::i18n::validation::name_length(
                    crate::constants::MAX_MEMORY_NAME_LEN,
                )))
            } else {
                Ok(())
            };
        assert!(resultado.is_err());
    }

    #[test]
    fn remember_response_merged_into_memory_id_some_serializes_integer() {
        let resp = RememberResponse {
            memory_id: 10,
            name: "mem-mergeada".to_string(),
            namespace: "global".to_string(),
            action: "updated".to_string(),
            operation: "updated".to_string(),
            version: 3,
            extraction_method: None,
            entities_persisted: 0,
            relationships_persisted: 0,
            relationships_truncated: false,
            chunks_created: 1,
            chunks_persisted: 0,
            urls_persisted: 0,
            merged_into_memory_id: Some(7),
            warnings: vec![],
            created_at: 0,
            created_at_iso: "1970-01-01T00:00:00Z".to_string(),
            elapsed_ms: 0,
            name_was_normalized: false,
            original_name: None,
            backend_invoked: None,
        };

        let json = serde_json::to_value(&resp).expect("serialization failed");
        assert_eq!(json["merged_into_memory_id"], 7);
    }

    #[test]
    fn remember_response_urls_persisted_serializes_field() {
        // v1.0.24 P0-2: garante que urls_persisted aparece no JSON e aceita valor > 0.
        let resp = RememberResponse {
            memory_id: 3,
            name: "mem-com-urls".to_string(),
            namespace: "global".to_string(),
            action: "created".to_string(),
            operation: "created".to_string(),
            version: 1,
            entities_persisted: 0,
            relationships_persisted: 0,
            relationships_truncated: false,
            chunks_created: 1,
            chunks_persisted: 0,
            urls_persisted: 3,
            extraction_method: Some("regex-only".to_string()),
            merged_into_memory_id: None,
            warnings: vec![],
            created_at: 0,
            created_at_iso: "1970-01-01T00:00:00Z".to_string(),
            elapsed_ms: 0,
            name_was_normalized: false,
            original_name: None,
            backend_invoked: None,
        };
        let json = serde_json::to_value(&resp).expect("serialization failed");
        assert_eq!(json["urls_persisted"], 3);
    }

    #[test]
    fn empty_name_after_normalization_returns_specific_message() {
        // P0-4 regression: name consisting only of hyphens normalizes to empty string;
        // must produce a distinct error message, not the "too long" message.
        use crate::errors::AppError;
        let normalized = "---".to_lowercase().replace(['_', ' '], "-");
        let normalized = normalized.trim_matches('-').to_string();
        let resultado: Result<(), AppError> = if normalized.is_empty() {
            Err(AppError::Validation(
                "name cannot be empty after normalization (input was blank or contained only hyphens/underscores/spaces)".to_string(),
            ))
        } else {
            Ok(())
        };
        assert!(resultado.is_err());
        if let Err(AppError::Validation(msg)) = resultado {
            assert!(
                msg.contains("empty after normalization"),
                "mensagem deve mencionar 'empty after normalization', obteve: {msg}"
            );
        }
    }

    #[test]
    fn name_only_underscores_after_normalization_returns_specific_message() {
        // P0-4 regression: name consisting only of underscores normalizes to empty string.
        use crate::errors::AppError;
        let normalized = "___".to_lowercase().replace(['_', ' '], "-");
        let normalized = normalized.trim_matches('-').to_string();
        assert!(
            normalized.is_empty(),
            "underscores devem normalizar para string vazia"
        );
        let resultado: Result<(), AppError> = if normalized.is_empty() {
            Err(AppError::Validation(
                "name cannot be empty after normalization (input was blank or contained only hyphens/underscores/spaces)".to_string(),
            ))
        } else {
            Ok(())
        };
        assert!(resultado.is_err());
        if let Err(AppError::Validation(msg)) = resultado {
            assert!(
                msg.contains("empty after normalization"),
                "mensagem deve mencionar 'empty after normalization', obteve: {msg}"
            );
        }
    }

    #[test]
    fn remember_response_relationships_truncated_serializes_field() {
        // P1-D: garante que relationships_truncated aparece no JSON como bool.
        let resp_false = RememberResponse {
            memory_id: 1,
            name: "test".to_string(),
            namespace: "global".to_string(),
            action: "created".to_string(),
            operation: "created".to_string(),
            version: 1,
            entities_persisted: 2,
            relationships_persisted: 1,
            relationships_truncated: false,
            chunks_created: 1,
            chunks_persisted: 0,
            urls_persisted: 0,
            extraction_method: None,
            merged_into_memory_id: None,
            warnings: vec![],
            created_at: 0,
            created_at_iso: "1970-01-01T00:00:00Z".to_string(),
            elapsed_ms: 0,
            name_was_normalized: false,
            original_name: None,
            backend_invoked: None,
        };
        let json_false = serde_json::to_value(&resp_false).expect("serialization failed");
        assert_eq!(json_false["relationships_truncated"], false);

        let resp_true = RememberResponse {
            relationships_truncated: true,
            ..resp_false
        };
        let json_true = serde_json::to_value(&resp_true).expect("serialization failed");
        assert_eq!(json_true["relationships_truncated"], true);
    }

    // GAP-08: body-preservation predicate tests.
    // Verifies the decision logic that determines whether an existing body should
    // be kept instead of overwritten with an empty incoming body during --force-merge.

    /// Returns `true` when the existing body should be preserved.
    ///
    /// Mirrors the `body_will_be_preserved` expression in `run()` so the logic
    /// is testable without a real database connection.
    fn should_preserve_body(force_merge: bool, raw_body_is_empty: bool, clear_body: bool) -> bool {
        force_merge && raw_body_is_empty && !clear_body
    }

    #[test]
    fn gap08_empty_body_force_merge_no_clear_body_preserves() {
        // Caller passes no body with --force-merge but without --clear-body.
        // The existing body in the DB must be kept.
        assert!(
            should_preserve_body(true, true, false),
            "empty body + force-merge + no clear-body should trigger preservation"
        );
    }

    #[test]
    fn gap08_empty_body_force_merge_with_clear_body_does_not_preserve() {
        // Caller explicitly passes --clear-body; intentional wipe is honoured.
        assert!(
            !should_preserve_body(true, true, true),
            "--clear-body must bypass preservation"
        );
    }

    #[test]
    fn gap08_non_empty_body_force_merge_does_not_preserve() {
        // Caller provides a real body; it must overwrite the existing one.
        assert!(
            !should_preserve_body(true, false, false),
            "non-empty body must overwrite, not preserve"
        );
    }

    #[test]
    fn gap08_empty_body_no_force_merge_does_not_preserve() {
        // Without --force-merge the path is a fresh create; no preservation needed.
        assert!(
            !should_preserve_body(false, true, false),
            "no --force-merge means no preservation logic applies"
        );
    }
}