trusty-memory

A place to put what was learned

An assistant that forgets everything at the end of a session relearns your codebase every morning. trusty-memory is the store that stops that: an MCP server over a local vector index and a key-value store, where an agent writes what it worked out and recalls it by meaning later.

Memories are organised into named palaces — one per project — with rooms and wings inside them. The naming is deliberate: a palace is anchored to a real project directory, so the memories for one repository can never quietly bleed into another's recall. Work outside any project and there is a single personal palace for the notes that belong to you rather than to a codebase.

Recall that does not need the words

Recall is hybrid — lexical BM25 alongside vector similarity — so a query finds a memory that said the same thing differently. Embeddings are computed on the machine, and both the vector index and the metadata store are ordinary local files. Nothing is shipped to a hosted memory service.

• memory_remember and memory_recall are the whole day-to-day surface; memory_recall_deep trades latency for reach when the fast lane comes up short.
• A knowledge-graph layer stores subject/predicate/object triples next to the prose, so structured facts can be asserted and queried directly rather than fished back out of a paragraph.
• A chat-session store keeps conversation turns verbatim, bypassing the signal filters that apply to ordinary memories — a transcript is not a fact and should not be deduplicated like one.
• memory_recall takes an optional min_score floor, so a caller can ask for only the hits that clear a relevance bar instead of filtering the response itself.

The dream cycle

A memory store that only ever accumulates degrades into a landfill. trusty-memory runs a consolidation pass — the dream cycle — that merges near-duplicates, prunes what has gone stale, and, when an inference backend is configured, summarises a room's older facts into canonical entries and links the originals to their replacement so the lineage stays traceable.

Task drawers — goals, milestones, checkpoints an application must re-derive across sessions — are exempt. They are never evicted and never consolidated away, however old they get, until something deletes them explicitly.

Wiring it up

trusty-memory setup installs the background service, warms the embedding model cache, and patches the Claude settings files it finds with the right server entry. From then on the daemon is just there.

For a manual configuration, the canonical entry runs trusty-memory serve --stdio, which forwards every request over the daemon's Unix domain socket and returns its answers verbatim. The stdio process never opens the database itself, so it coexists safely with the daemon and with other clients. The daemon has no HTTP listener and no REST API — everything goes through that one socket. The admin dashboard that used to live here moved to trusty-console, which reads the same socket and serves it at /tools/memory/.