Skip to content

Learning Loop

The loop: act → observe → distill → persist → apply. Everything it learns is visible, attributed, and reversible.

1. Watch a lesson get born

  1. Open a project and give G-Rump a task that fails once before succeeding (e.g. a build with a missing generate step).
  2. After the run, a chat notice appears: Learning: saved 1 lesson… (toggle in Settings → Brain → Learning).
  3. Open the Learning panel (graduation cap in the right dock) → Lessons: the new lesson shows a 50% confidence bar (Laplace prior — no track record yet).

2. Watch it ride along and earn confidence

  1. Ask for a similar task. The lesson injects into the system prompt (## Learned Lessons block, top 5 by relevance × confidence).
  2. Success → the lesson's win count and bar go up. Say "that's wrong" instead → the run is amended in Outcomes and the lesson takes the loss.
  3. Lessons that keep losing auto-retire at <30% confidence after 5 rides.

3. Watch it propose a skill

  1. Once ≥3 lessons cluster around one workflow, reflection may emit a proposal — the Learning tab badges, and the Proposals tab shows a unified diff of the SKILL.md it wants to write, with rationale and the source lessons.
  2. Approve & Enable writes the skill and turns it on. Reject is remembered forever — it will never be re-proposed.
  3. There is no other path: the agent cannot write SKILL.md, SOUL.md, or MIND.md through file tools without an explicit approval prompt.

4. Watch the daemon get pickier

  1. Queue goals via the add_goal tool (or vault Goals/). The daemon scores priority + 2 × success-rate per task type, parks types it keeps failing as needs-attention, and reflects after every goal.

Manual controls: reflect tool for an on-demand pass · pin/retire/edit any lesson in the panel · the whole loop switches off in Settings → Brain → Learning.