Recall

Memory for agents. Episodic turns, semantic embeddings, working sets, and a scoped KV — all in one schema, all hit through agent-memory.

Tables

Table	Purpose
recall.episodic	Ordered turns (kind, text, provenance, scope).
recall.semantic	Embedded chunks with a Firestore vector search column.
recall.working_sets	Named, bounded sets the agent actively reasons over.
recall.kv	Scoped key/value (per-agent, per-scope).

Scopes

Every write carries scope text[] (e.g. ['self'], ['team:triage'], ['org']). Reads intersect the caller's authorized scopes. A leaked device key for agent A can never read B's ['self'] rows.

Similarity search

recall.semantic uses Firestore vector search with an ivfflat index. The Cloud Function exposes /mem/query:

{
  "agent_id": "agent-hello",
  "q": "what did I promise the vendor?",
  "k": 8,
  "scope": ["self", "team:sales"]
}

Firestore vector search returns cosine-ranked chunks; provenance rides back so the agent can cite.

Episodic turns

Turns are append-only and carry provenance (e.g. inbound-email:<msg-id>, slack:<ts>) so downstream summaries trace back to primary sources.

Routes

Method	Path	Purpose
POST	/mem/write	Append an episode, semantic chunk, or KV row.
POST	/mem/query	Similarity or exact-match read.
POST	/mem/forget	Soft-delete by id or provenance.
POST	/set/put	Update a working set.
POST	/kv/get	Scoped KV fetch.
POST	/kv/put	Scoped KV write.

Retention

agentpack_recall_episodic_retention (03:37 UTC daily) trims episodic rows past recall_episodic_retention_days (default 180). Semantic and KV are retained until explicit forget.