Recoup — Capture Lesson (compounding memory)

The box gets smarter per use by writing down what it learns. The discipline: research broadly, write once, never duplicate, make the store discoverable.

When this fires

The user signals a win or "remember this", or a non-trivial problem was just solved (a fix, a workaround, a judgment call). Don't fire for trivial one-liners or generic chat — a learning must be reusable.

Procedure

1. Gather (may read; don't write yet). Reconstruct: Problem (what went wrong, one sentence) · Context (where it shows up) · Root cause (why, not just the symptom) · Solution (concretely — commands, the exact field, the decision) · Prevention (what to do first next time). Unknown → ask, don't invent.

2. Dedup before create. List existing learnings in the target store; score overlap against each on five dimensions (problem / root cause / solution / referenced files / prevention). High overlap → update the existing file, don't create a near-duplicate ("two docs on the same problem drift apart").

3. Write exactly one file, filed by primary subject:

   • tied to one artist → artists/{slug}/learnings/{slug}.md
   • tied to one deal → deals/{id}/learnings/{slug}.md
   • cross-cutting (a platform/API/process lesson) → root learnings/{slug}.md

   Frontmatter (problem_type, component, tags, date) + body (Problem / Context / Root cause / Solution / Prevention). Commit learn: <title>.

4. Discoverability check. A learning nobody reads is dead weight — confirm a fresh agent would find the store (does the workspace README/RECOUP.md mention learnings/? add the smallest natural pointer if not).

Guardrails

• One file per capture — research may range; the write is single + deterministic.
• Update, don't duplicate — high overlap edits the existing learning.
• File by subject, not by skill — a voice lesson lives under the artist, even if a content task surfaced it.
• Reusable only; never fabricate a root cause or solution.