@msm-core/learning

Auto-improvement for durable agents. Close the loop from outcome to better behaviour.

plan runs (verify verdicts + escalations) ──▶ recordExperience
   ▲                                            (goal, plan, outcome, failure, human feedback)
   │ seed the next decompose                              │
   │                                                reflect (brain → lessons)
 retrieveLessons + formatLessons  ◀─────────  winning approaches / anti-patterns
   (relevant, confidence-gated)                 (keyed by goal meaning, confidence-gated)

The verify-and-adapt verdicts an agent produces (see @msm-core/planner) are exactly the training signal this consumes. Storage- and brain-agnostic — inject a LearningStore and a judge.

Pieces

• recordExperience(store, exp) — capture a goal pursuit (call it on a job's terminal state).
• reflect(store, judge, opts) — distil recent experiences into lessons; persists them. Best-effort (too few experiences / brain outage / junk → no lessons, never throws).
• retrieveLessons(store, goal, opts) — the lessons most relevant to a goal, confidence-gated.
• formatLessons(lessons) — render them as a prompt fragment to seed a decompose/replan.
• InMemoryLearningStore — a reference store (word-overlap similarity) for tests + light use; back it with a vector index in production so lessons travel with the agent.

Safe by design

• Advisory, not authoritative — lessons seed the prompt; the brain still decides. A bad lesson nudges, it never forces.
• Confidence-gated + mergeable — a lesson the agent keeps re-learning gains support and confidence; a one-off stays weak and is filtered out of retrieval.
• Sovereign — experiences and lessons live in the store you provide (e.g. the agent's own database), never leave it.

Zero runtime dependencies.