Search past Claude Code sessions (episodes), memory notes (memories), and per-session digests (summaries, written by the nightly reflect pass).

Hybrid full-text + semantic search. scope: "all" | "episodes" | "memories" | "summaries" ("all" spans all three). Filters: project matches path segments of the session dir (e.g. "mobile" => ~/src/mobile); since/until are "YYYY-MM-DD" or "Nd" (N days ago), inclusive. Returns compact hits: kind, id, date, project, snippet, session_id (episodes and summaries) or name (memories), and a score in (0, 1] (recency- and kind-weighted, normalized RRF); hits are score-descending. Episode/memory ids are stable content-derived strings (ep_…/mem_…) — safe to persist; summary ids (sum_…) are ephemeral. Use get_episode(id) for full text.

Fetch full stored text + metadata for one recall hit.

id is the string from a recall hit; its ep_/mem_/sum_ prefix selects the kind (episode → the USER/ASSISTANT exchange; memory → body + path; summary → a per-session digest written by reflect). ep_/mem_ ids are stable across re-ingests and rebuilds; sum_ ids are ephemeral (pair recall → get_episode within one session). Text capped at 8k chars with a truncation marker.

Persist a new memory: write memories/.md, commit, index — immediately recallable.

fact: the durable fact (short markdown, first sentence becomes the description). type: user|feedback|project|reference. name: optional kebab-slug filename. session/project: provenance trailers ($CLAUDE_CODE_SESSION_ID from Bash env / workspace dir); omitted when empty. If a near-duplicate memory exists it is returned instead of writing; force=true writes anyway. On success returns a stable mem_… id — pair recall → get_episode(id) to read it back.