Deep memory

Deep memory paths are slower than ordinary recall because they use planning, multiple searches, reflection, synthesis, and sometimes revision. They are for moments where a single vector search is not enough.

Deep recall

Deep recall plans several search queries from the user's question, searches memory, reflects on whether the results are sufficient, and synthesizes a curated result set. The MCP renderer shows the trace: plan, searches, reflection turns, synthesis, summary, and memory IDs.

Use it when the query is ambiguous, cross-session, or depends on patterns rather than a single known phrase.

Deep remember

Deep remember searches for related memories before writing. It can skip a duplicate, refine a related record, or store a new one. This prevents long-lived memory from filling with repeated versions of the same decision.

The default MCP mode is auto: semantic/foundational content tends to go deep, while time-stamped episodic events can stay shallow.

Deep compose

The MCP compose tool uses memory to produce a deliverable. It builds a section plan, retrieves memories for each section, drafts, critiques, revises, and can write an episodic anchor for the composed work.

Use compose for handoff notes, architecture summaries, implementation plans, and decision records where the output must reflect what the user has already done across sessions.