curate-learning

name	curate-learning
description	Review captured learning, compare it with existing knowledge and Engram, validate evidence, reject noise, resolve duplicates or conflicts, and approve an exact destination. Curating never writes Skills or AGENTS.md.
license	MIT
metadata	{"author":"RoyoTech","version":"3.0.0","mcp_profile":"agent"}

Curate Learning

Workflow

1. Load the candidate with learning_get.
2. Search related learnings and Engram.
3. Identify relation: new, duplicate, extends, supersedes, contradicts, narrows, related.
4. Validate using tests, diffs, documentation, reproduction, or independent cases.
5. Decide exactly one:
   • reject;
   • needs evidence;
   • merge;
   • approve project/shared knowledge;
   • approve new/update Skill;
   • approve AGENTS rule;
   • approve regression test.
6. Define exact destination, acceptance checks and rollback condition.
7. Call learning_curate.

Destination rules

• project facts and status → project knowledge or Engram;
• repeatable procedure → Skill;
• short mandatory invariant → AGENTS.md;
• evidence too weak → needs evidence;
• duplicate → merge/reject;
• preference → personal/project unless user explicitly universalizes it.

A universal rule requires stronger evidence than a project note.

Do not publish.