Record a Lesson Learned

After fixing a bug, resolving an incident, or discovering an important pattern, record it for future sessions.

Steps

1. Classify the domain. Determine which memory file this belongs to:

   • Pipeline/cron/Lambda/workers → memory/pipeline-ops.md
   • Data quality/known bugs → memory/data-quality.md
   • UI/frontend/navigation → memory/ui-navigation.md
   • MCP server/CLI → memory/mcp-server.md
   • Cross-cutting/general → memory/lessons-learned.md
   • If it's a critical safety rule → also add to MEMORY.md (keep under 200 lines)

   Before picking a doc tier at all, ask whether this should be a CHECK. A doc only works if the next person reads it at the moment it applies, and that is a weak bet: measured 2026-08-21, three of four findings in one session were classes where the doc already existed and the thing recurred anyway (#4163, #4190). If you can name the file or symbol that must hold the property, prefer a sweeping test, a detector that files what it finds, or a constructor that throws — and let the doc be one line pointing at it. Run the negative control first (invariants/tests-that-are-not-guards.md): delete the guarded line, watch it go red, restore it. Keep prose for lessons about judgment, which cannot be asserted.

   If it rises to an invariant — a rule whose violation already cost real damage — pick the tier deliberately:

   • Applies regardless of what you're working on → CLAUDE.md (budget: ~5,500 words by wc -w; over it, demote something to invariants/ first).
   • Fires only when someone touches a particular subsystem → a new or existing .claude/docs/invariants/<name>.md, plus a one-line trigger entry in CLAUDE.md's routing table. If you can name the file or subsystem that triggers the rule, it belongs here — this is the common case.

2. Check for contradictions. Read the target memory file and check:

   • Does this contradict an existing entry? → Update the old entry, don't duplicate.
   • Does this make an existing entry obsolete? → Remove or update it.
   • Is there already an entry for this? → Strengthen it with new detail, don't duplicate.

3. Write the entry. Format:

   - **Short description (YYYY-MM-DD):** What happened, what the fix was, and the general rule to follow going forward.
   

   • Always include the date.
   • State the rule (what to do), not just the story (what happened).
   • Keep it to 1-2 lines. Link to files/docs if details are needed.

4. Check staleness nearby. While editing, scan neighboring entries:

   • Any entry with a date older than 30 days → verify it's still accurate or mark (verify).
   • Any entry referencing a file path → quick-check the file still exists.
   • Any entry with stats/counts → mark as (snapshot, may be stale) if not already marked.

5. Report. Tell the user what was added, what was updated, and any stale entries found.

Examples

Good entry:

- **Lambda timeout on large books (2026-03-13):** Books with >500 pages can exceed Lambda 15min timeout. Split into chunks of 400 pages max. See `src/lib/lambda-queue.ts:128`.

Bad entry (no date, no rule):

- Fixed the Lambda timeout issue.

When NOT to use this

• Trivial fixes (typos, one-line changes) — not worth recording
• Session-specific context — use handoffs instead
• Speculative conclusions from reading one file — verify first