Struct MemoryManager 

pub struct MemoryManager { /* private fields */ }

Agent memory manager.

Stores and retrieves memory entries using the file-based StateStore. Supports embedding-based vector search via an in-memory TF-IDF index that is rebuilt on startup.

Implementations

impl MemoryManager

pub async fn total_entries(&self) -> usize

Returns total entries across all memory types (from disk).

pub async fn rebuild_index(&self) -> Result<()>

Rebuild the vector index from all stored memories.

Call once at startup to populate the in-memory index from persisted memory entries.

pub async fn save_index_snapshot(&self) -> Result<()>

Save the current vector index to disk as a snapshot.

pub async fn load_index_snapshot(&self) -> Result<usize>

Load a previously saved vector index snapshot from disk.

pub async fn remember(&self, entry: MemoryEntry) -> Result<String>

Store a memory entry. Returns the entry ID.

Also computes and stores the entry’s text vector in the in-memory index for future semantic search.

pub async fn get( &self, id: &str, memory_type: MemoryType, ) -> Result<Option<MemoryEntry>>

Retrieve a single memory by ID.

pub async fn forget(&self, id: &str, memory_type: MemoryType) -> Result<bool>

Delete a memory entry.

pub async fn list( &self, memory_type: MemoryType, limit: usize, ) -> Result<Vec<MemoryEntry>>

List memories of a given type, most recent first.

pub async fn search( &self, query: &str, memory_type: Option<MemoryType>, limit: usize, ) -> Result<Vec<MemoryEntry>>

Search memories by semantic similarity (vector search).

Falls back to keyword search when the vector index is empty or yields no results above the similarity threshold.

pub async fn recall(&self, query: &str) -> Result<Vec<MemoryEntry>>

Recall relevant memories for a new session.

Combines recent conversation summaries, session summaries, and keyword-matched facts/episodes.

pub fn blend_into_prompt( &self, memories: &[MemoryEntry], system_prompt: &str, ) -> String

Blend recalled memories into the system prompt.

pub async fn summarize_session( &self, session: &Session, ) -> Result<Option<String>>

Create a session summary memory entry from a completed session.

This does NOT use LLM — it records key metadata from the session as a structured memory entry for future reference.

pub async fn is_duplicate(&self, content: &str) -> bool

Check if a memory entry with identical content already exists.

Uses a fast hash comparison against the in-memory vector index.

pub async fn remember_unique( &self, entry: MemoryEntry, ) -> Result<Option<String>>

Store a memory entry only if no duplicate content exists.

Returns the entry ID if stored, or None if duplicate.

impl MemoryManager

pub async fn semantic_search( &self, query: &str, memory_type: Option<MemoryType>, limit: usize, hnsw_index: &HnswMemoryIndex, ) -> Result<Vec<SemanticHit>>

Semantic search using HNSW index.

Unlike search() which uses brute-force cosine similarity over the in-memory HashMap, semantic_search() uses the HNSW approximate nearest neighbor index for sub-linear time complexity.

This is the preferred search method when the HNSW index is available and populated with dense vectors.

Arguments

• query — Search query text.
• memory_type — Optional filter by memory type.
• limit — Maximum results to return.
• hnsw_index — The HNSW index to search against.

Returns

A list of SemanticHit with entry and similarity score.

pub async fn rebuild_hnsw_index( &self, hnsw_index: &HnswMemoryIndex, ) -> Result<usize>

Rebuild the HNSW index from all stored memories.

Call this at startup or after bulk operations.

impl MemoryManager

pub fn new(state_store: Arc<StateStore>) -> Self

Create a new MemoryManager.

pub fn set_git_layer(&mut self, gl: Arc<GitLayer>)

Attach a git layer for version-controlled saves.

pub fn for_space(space_dir: PathBuf) -> Self

Create a Space-scoped MemoryManager.

Each Space gets its own StateStore under the given directory, providing natural memory isolation between Spaces.

pub fn set_hnsw_index(&self, index: Arc<HnswMemoryIndex>)

Attach an HNSW index for fast semantic search.

Once attached, remember() and forget() automatically keep the HNSW index in sync with the state store.

pub fn with_max_recall(self, n: usize) -> Self

Set max memories returned by recall.

pub fn with_config(self, config: &MemoryConfig) -> Self

Apply MemoryConfig settings.

pub fn vector_index_size(&self) -> usize

Returns the number of entries in the vector index.

pub fn effective_importance(entry: &MemoryEntry) -> f32

Compute effective importance of a memory entry.

Effective importance = base_importance * (1 + log(1 + access_count)) Memories accessed frequently get a boost.

pub async fn curate(&self, budget: &MemoryBudget) -> Result<CurationReport>

Curate memories: identify candidates for removal based on budget.

Returns a report of how many entries would be pruned per type.

pub fn spawn_curation_task(self: &Arc<Self>, budget: MemoryBudget)

Spawn a background curation task.

Returns immediately; curation runs asynchronously.

Trait Implementations

impl Debug for MemoryManager

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.