Working Memory

The agent's immediate context — recent observations, active conversations, and current tool results. Fastest access, limited capacity, and subject to eviction.

Schema

Indexes

• idx_working_session on session_id
• idx_working_timestamp on timestamp
• idx_working_source on source

FTS5

Full-text search is maintained via fts_working (content-column FTS5) with automatic triggers for INSERT, UPDATE, and DELETE.

Vector Search

Embeddings are stored separately in vec_episodes (sqlite-vec, int8[384]). They are generated at recall time and are not a column in the working_memory table.

Eviction Policy

When Working Memory reaches capacity, entries are evicted by a composite score:

1. Importance weighting: Higher importance = lower eviction priority
2. Recency: Recently created or recalled entries are protected
3. Recall frequency: Entries with higher recall_count are protected
4. Composite score: score = importance * recency_factor * recall_factor

Consolidation Path

flowchart LR
  A[New Entry] -->|insert| WM[Working Memory]
  WM -->|TTL/2 age trigger| C{Consolidation}
  C -->|group by source| S[Summarize via LLM/AAAK]
  S -->|promote| EM[Episodic Memory]
  C -->|evict originals| DEL[Evicted]

API Reference

# Store a memory
mem.remember(content, importance=0.5, source="user")

# Recall memories (FTS5 + vector hybrid)
results = mem.recall(query, top_k=10)

# Update a memory
mem.update(id, content=None, importance=None)

# Delete a single memory
mem.forget(id)

Performance