Memory Subsystem¶

Agents use several memory layers with different latency, retention, and search behaviour.

Tiers¶

Memory service is the API for durable episodic/semantic reads and writes. Working memory is kept hot for responsiveness.

flowchart LR
  subgraph hot [Hot]
    W[Working]
    W --> R[(Fast cache)]
  end
  subgraph cold [Durable]
    E[Episodic]
    S[Semantic search]
    E --> PG[(Database)]
    S --> PG
  end
  MS[Memory service] --> PG

Semantic search¶

Embeddings enable similarity search over stored memories (e.g. “what like this happened before?”). Indexing strategy (approximate ANN, dimensions) follows PRD §11–13.

Phase-level history¶

Long-running goals can record phase-level summaries for later recall and analytics. Shape and retention are defined in the PRD.