mnemo

Mnemo — persistent memory

You have access to a persistent memory layer via these MCP tools:

• mnemo_recall(query, k?, scope?) — semantic search across stored memories
• mnemo_remember(content, scope?, tags?) — capture a new memory
• mnemo_forget(id) — delete a memory
• mnemo_list(scope?, limit?) — browse recent memories
• mnemo_stats() — show memory engine status

When to use each

mnemo_recall — call this BEFORE answering anything where prior context might help:

• "what's our convention for X?"
• "how did we decide to handle Y?"
• "why is Z structured this way?"
• whenever you're starting a task in an unfamiliar area of the codebase

Use natural-language queries. Don't try to construct keyword searches — the index is semantic.

mnemo_remember — call this when:

• The user explicitly asks ("remember this", "save this for later")
• You make or affirm a non-obvious decision (capture it as scope: "project")
• The user states a personal preference that applies across projects (capture as scope: "global")
• You finish a task and notice something worth not having to rediscover

Default scope is "project" (auto-tied to the current repo).

mnemo_forget — only when the user explicitly asks to forget, OR when you discover a memory contradicts current reality and should be replaced.

Examples

User: "we always use Vitest, never Jest" → mnemo_remember({ content: "we always use Vitest, never Jest", scope: "global" })

User: "what test framework do we use?" → mnemo_recall({ query: "test framework preference", k: 3 }) then synthesize the answer.

User: "forget that pnpm thing" → mnemo_list({ limit: 50 }) to find it, then mnemo_forget({ id: "..." }).

Discipline

• Don't capture things that are already in CLAUDE.md or other always-loaded files (Mnemo is for the long tail that those files can't carry)
• Don't capture transient task state — capture lasting facts
• Prefer global scope for cross-project preferences; default to project scope otherwise
• When recall returns nothing useful, say so plainly — don't guess or hallucinate from low-similarity hits