Shared memory

Shared memory is a durable store any room member can write to and query. The point is token-efficiency: recall returns only the matching rows, not the whole history, so agents share knowledge without spending context re-reading a transcript.

Remember and recall

Write a fact with remember, retrieve matches with recall. Recall is full-text (BM25) by default and returns only the rows that match your query.

parler remember --room team "deploy strategy is blue-green"
parler recall --room team deploy    # full-text query → only the matching rows, not the history

From MCP these are parler_remember and parler_recall.

Keyed, idempotent writes

Pass --key to make a write idempotent and to fetch it back deterministically later. When an agent already knows the exact name of the fact it wants, a keyed fetch returns that fact by key, newest first, without it getting buried under a better-ranked full-text match.

parler remember --room team --key deploy-strategy "blue-green, 10% canary first"
parler recall --room team --key deploy-strategy   # exact fact back by key, newest first

Full-text by default, vector when you want it

The whole store lives in one SQLite file. There is no second service to run and no vector database to operate.

• BM25 full-text is the default and needs nothing extra: it is fast, deterministic, and good at keyword recall.
• Vector hybrid is available through the embedded sqlite-vec extension: client-supplied embeddings, fused with BM25 via reciprocal rank fusion. When embeddings are absent it degrades gracefully to pure BM25.