| name | memnos-memory |
| description | Long-term memory via memnos. Use when the user references past work, prior decisions, preferences, people, projects, or anything from earlier conversations — and after giving answers that contain decisions, identifiers, or outcomes worth keeping. |
memnos — long-term memory
You have persistent, governed long-term memory through the memnos MCP tools
(recall, recall_wide, remember, reconcile_claim, get_entity, get_provenance).
Claude Desktop has no automatic memory hooks, so YOU are responsible for using these tools
consistently. Follow these rules:
Recall — before answering
- When the user references past conversations, prior decisions, preferences, ongoing
projects, or people/things not introduced in this session, call
recall with a focused
query before answering. If nothing relevant returns, call recall_wide (searches all
namespaces this token can read).
- Prefer recalled facts over guessing. If a recalled fact conflicts with what the user just
said, use
reconcile_claim and surface the discrepancy with the dates.
Remember — after answering
- After any answer that contains a decision, conclusion, identifier, or outcome (ticket
keys like ABC-123, PR/MR numbers, versions, URLs, chosen options, agreed plans), call
remember with a ONE-LINE summary of what was decided/produced — keep identifiers
verbatim, never paraphrase them away.
- Also
remember durable facts the user states about themselves, their projects, their
preferences, and their commitments. One fact per call, self-contained, with dates
resolved to absolute form.
- Do NOT store small talk, transient scratch work, secrets/credentials, or restatements of
things already remembered this session.
Notes
- Memories are namespace-scoped and access-controlled server-side; if a write/read is
denied, say so rather than retrying.
- If the tools report the memnos server is not running, tell the user to run
memnos start — do not silently continue without memory.
