learn

Learn — Sweep Raw Memory, Crystallize Explicit Lessons

Two jobs, in order: (1) mechanically true up the chunk index so every conversation and doc is searchable memory, and (2) write a vault note for every explicit lesson this session — user corrections, explicit save-requests, presented conclusions that were later overturned, and confirmed approaches (user-praised or self-validated). Nothing else.

Raw event memory is AUTOMATIC. engram ingest chunks and embeds session transcripts and markdown itself — no summaries, no episode notes, no arc detection. The agent's only writing job is crystallizing EXPLICIT lessons. Do not reconstruct the old episode workflow (engram transcript, engram learn episode) — it is gone.

Mid-cycle capture (fast path). When you fire at a CORRECTION moment mid-task — a review, a failing check, or the user just rejected your approach or named a different convention — this is a focused single-note capture, NOT the open/close of a cycle. SKIP Step 1 (sweep) and Step 1.5 (vocab); go straight to Step 2 and crystallize the ONE confirmed correction (hand off to write-memory as always). Do NOT run engram ingest --auto mid-task — the sweep is a cycle-boundary job, it is wasteful here and can block on a large corpus, and losing the capture to a hung sweep is the failure mode. The closing learn still sweeps.

Step 1 — Sweep (first; the closing learn always sweeps)

engram ingest --auto

That's it. The binary stats every known source (repo markdown, .claude dirs, all session transcripts), re-chunks and re-embeds only what changed — within one source, existing chunks are never deleted (append-only history). Across sources, byte-identical content is deduplicated: only one canonical copy is indexed, and a duplicate's index is removed only once its retained twin is verified to cover its records (not on hash-match alone) — see engram prune --duplicates for the retroactive cleanup mode. Unchanged corpus → returns in seconds. Report the one-line tally it prints (or "memory index up to date").