Memory and Learning — A Toolkit That Gets Better
Skills That Evolve, and Learning Intake
Improving the Toolkit From Evidence
Skills are not frozen. /review:evolve-skills mines two signals from your own repository — misalignments logged by agents (proposing IMPROVE) and clusters of recurring workflow/pattern memories (proposing CREATE) — and writes skill proposals.
The rule that keeps this honest is a differential eval: a proposed skill must beat the baseline (with_skill versus without_skill or the old version) before it is applied. Evidence of a gap comes first; the eval comes after; nothing is applied on vibes.
New in 6.2: attribution uses IDF weighting, a distinct-token gate and a confidence margin. Near-ties are reported as ambiguous instead of being attributed, and an explicit "none attributable to the agents" section counts as caught. IMPROVE scores from this version are not comparable with earlier ones.
GitHub-Based Learning Intake
v6.2 also includes a learning-intake workflow that can send a captured learning to a GitHub Issues receiver. The safe defaults matter:
- Disabled by default. Installing or updating never turns delivery on and never contacts GitHub
- GitHub Issues only; each project sets its own receiver as
OWNER/REPO - Initialization creates disabled defaults and cannot enable delivery or grant approval
A disabled local success is not proof of delivery. When the workflow reports a result, read the exit code and the bounded result rather than assuming a send happened.
Before a proposed skill change from /review:evolve-skills is applied, what must be true?