remember-user-fact

Remember a user fact

When you learn something about the user that would be useful next session, save it now — don't defer to end-of-session. Small, atomic saves are easier to retrieve and update than one big dump later.

When to invoke

• Explicit request: "remember that…", "save this for next time", "store in memory".
• Correction: the user tells you to change your behavior ("don't do X", "always do Y", "stop assuming Z"). These are high-value — save them so the next session doesn't repeat the mistake.
• Confirmed preference: the user states a stable preference with reasoning, not just an offhand comment. Reasoning is the signal that it's durable.
• Durable project context: you learn a client name, ongoing initiative, long-running constraint, or recurring stakeholder that will matter beyond this session.
• Identity/role details: role changes, new responsibilities, location changes, tooling switches.

Do not invoke for:

• Transient conversation state (current file, current task).
• Information already in code, git, or project docs — memory is for facts you cannot derive from the filesystem.
• One-off frustrations or snap opinions the user may not hold tomorrow.
• Secrets, credentials, tokens, or anything sensitive.

How to run it

1. Load the config — read .claude/memory-config.md in the workspace. That file names the backend, the exact MCP tool to call for "add", and the scope parameters (index/namespace/project_id/etc.) for the chosen context. If the file is missing, stop and ask the user to install one (copy the plugin's templates/memory-config.example.md).
2. Pick the context — apply the deduction rule from CONTEXT.md. If unsure, ask one short question ("save this to personal or work memory?").
3. Call the configured add tool with the context's scope parameters. Construct the record according to the schema in memory-config.md; at minimum:
   • A self-contained natural-language statement of the fact, written from the user's perspective as if they were telling a future assistant. Example: "I prefer pnpm over npm for new Node projects — npm's lockfile churn has burned me twice." Include the why if it was given; future-you uses the why to judge edge cases.
   • An absolute YYYY-MM-DD date.
   • Optional metadata the config recommends (type=preference/correction/project-context, topic tags, etc.).
4. Before saving, check for duplicates — run a quick semantic search for the same topic using the backend's search tool. If a near-identical memory already exists, prefer an update over creating a second record. Duplicates pollute retrieval.
5. Confirm briefly — one short line: "Saved to personal memory." or "Saved to work memory." Do not quote the full memory back.

Quality bar

• One fact per memory. "Prefers pnpm; based in Jerusalem; works on X" is three memories, not one. Atomic memories are retrievable; compound ones are not.
• Write for retrieval, not for logs. The memory will be embedded and searched — use the keywords someone looking for this info would actually use.
• Include the why when you have it. "Prefers pnpm because npm's lockfile churn burned them twice" retrieves and reasons better than "prefers pnpm."