memory

name

memory

description

Long-term memory (vacuole): search, read, store, forget, sync. Use when (1) user asks about past decisions, preferences, people, todos, or dates — call memory_search or memory_recall then memory_get; (2) user says 'remember this' — memory_store(text, target?); (3) mark something to forget — memory_forget(text); (4) after bulk edits to MEMORY.md or memory/* — memory_sync. Workspace: MEMORY.md and memory/YYYY-MM-DD.md under agent dir (default ./.u or MEMORY_WORKSPACE). base_skill does raw read/write; this skill adds search (FTS5, optional hybrid embedding), Store/Forgotten sections, and daily logs.

Memory

The agent’s long-term memory lives in the workspace: MEMORY.md and memory/YYYY-MM-DD.md. base_skill can read/write any file; this skill adds search (FTS5, optional hybrid embedding), memory_store (## Store or daily log), memory_forget (## Forgotten), and memory_sync (rebuild index). What’s on disk is the source of truth for recall.

When to use which tool

Goal	Tool
Answer about past / preferences / who / when	memory_search(query, maxResults?) or memory_recall(…) then memory_get(path, from?, lines?) as needed
Persist “remember this”	memory_store(text, target?) — appends to MEMORY.md ## Store or today’s memory/YYYY-MM-DD.md
Mark for forgetting (review later)	memory_forget(text) — appends to MEMORY.md ## Forgotten; does not delete
Refresh search after bulk edits	memory_sync — rebuilds FTS5 index (Node 22+); optional embedding and session transcript

Workflow

1. Recall: memory_search or memory_recall → use memory_get to read full snippets.
2. Store: memory_store(text) (longterm) or memory_store(text, "daily") for today’s log.
3. Forget: memory_forget(text) to append to ## Forgotten for human review.
4. After many file changes and Node 22+: memory_sync to rebuild the index.

Layout and env

Workspace root: default ./.u or MEMORY_WORKSPACE (in monoU often the agent dir, e.g. .first_paramecium). Paths: MEMORY.md (## Store, ## Forgotten), memory/YYYY-MM-DD.md, optional memory/index.sqlite.
Full layout and env: references/layout.md.

Short-term vs long-term

• Same-session multi-turn: Handled by runtime state.messages; no tools needed.
• Across sessions: Use memory_store or write to persist; then memory_search to recall.

See references/short-term-memory.md.

Pre-compaction flush

Before context compaction, you can run a silent turn so the model writes lasting notes to disk. From @monou/agent-from-dir: runMemoryFlushTurn(session, state, config, streamFn, { prompt?: string }). Default prompt: "Session nearing compaction. Write any lasting notes to MEMORY.md or memory/YYYY-MM-DD.md now. Reply with NO_REPLY if nothing to store."

Memory vs knowledge

• memory: What happened, decisions, preferences, people, todos, dates. Use for “what did we decide / my preference.”
• knowledge: Reference material, docs, FAQ (KNOWLEDGE.md, knowledge/*). Use for “according to the docs.”