reflect

reflect

You are running the brain's consolidation pass. The brain lives at ~/src/brain: memories/ is the declarative store (markdown, git = source of truth), cache/brain.db is a disposable derived index. Work directly; be economical — in headless mode nobody is watching, so never ask questions and never wait.

Steps

1. Sync, then refresh the index. First, best-effort pull: uv --directory ~/src/brain run brain-sync pull — on failure (offline, no remote, or a cache/sync-conflict flag) note it in the report and continue; it never blocks the pass. Then uv --directory ~/src/brain run brain-ingest (incremental; near-instant when nothing changed).

2. Find the review window. Read ~/src/brain/cache/last-reflect (ISO timestamp). Missing file → default to the last 48 hours.

3. Pull the time slice. Query the index directly for episodes since then: sqlite3 ~/src/brain/cache/brain.db "SELECT ts, session_id, cwd, substr(text,1,600) FROM episodes WHERE ts > '<last>' AND is_sidechain = 0 ORDER BY ts". Read full exchanges only where the slice looks consequential (decisions, corrections, new facts) — not routine tool chatter.

4. Write session digests. Find digest-worthy sessions from the index with exactly this query — the strftime format string is load-bearing (episode ts is ISO-8601 with a T separator, so a bare datetime('now'), which uses a space, would break the lexical -1h comparison and let live sessions through). Do not improvise it:

   SELECT session_id, count(*) AS n, sum(length(text)) AS chars,
          substr(min(ts),1,7) AS ym, substr(min(ts),1,10) AS day, cwd
   FROM episodes
   WHERE is_sidechain = 0
   GROUP BY session_id
   HAVING (n >= 3 OR chars >= 15000)
      AND max(ts) < strftime('%Y-%m-%dT%H:%M:%S', 'now', '-1 hour')
   ORDER BY max(ts) DESC;
   

   Privacy denylist: drop any candidate whose cwd contains a substring listed in BRAIN_DIGEST_EXCLUDE (config env var; e.g. personal-finance projects) — those are never digested and never written to summaries/. For each remaining session, skip it if ~/src/brain/summaries/*/<session_id>.md already exists — unless its current episode count (n) is at least double the episodes: value in that file's frontmatter (a resumed session), in which case rewrite the digest in place at that exact glob-matched path — never recompute the YYYY-MM directory on a rewrite, or an old-month copy would survive and double-index the session. Take at most 5 per night, newest first; leftovers wait for the next night. For each selected session, read its episodes from the index (SELECT ts, substr(text,1,2500) FROM episodes WHERE session_id = '<id>' AND is_sidechain = 0 ORDER BY ts; pull more of an episode only when the slice is clearly consequential) and write summaries/YYYY-MM/<session_id>.md (new digests only — YYYY-MM from the first episode's ts); frontmatter session_id, date, project (basename of cwd), cwd, host (hostname -s), episodes (count), description (one line, ≤150 chars); body ≤ ~1,800 chars: what happened, decisions, facts learned, files touched, outcome. Same bar as memories: never secrets/credentials, no transcript quotes longer than a sentence.