| name | capture-learning |
| description | Capture a reusable lesson after a verified bug fix, user correction, non-obvious discovery, recurring failure, or successful procedure. Search first, then call royo-learn. Capturing never publishes or changes agent behavior. |
| license | MIT |
| metadata | {"author":"RoyoTech","version":"3.0.0","mcp_profile":"agent"} |
Capture Learning
Use when
- the user corrects the agent;
- a non-obvious bug is fixed;
- a failed approach reveals a prevention rule;
- a procedure succeeds repeatedly;
- a tool limitation is discovered;
- the same explanation is being repeated;
- a significant task ends with a reusable lesson.
Do not use for routine progress or a session dump.
Workflow
- Search
learning_search using the central terms.
- Search Engram through the same tool when useful.
- Decide whether there is an atomic, reusable lesson.
- Separate direct observation from interpretation.
- Gather only necessary evidence.
- Call
learning_capture.
- Do not curate or publish in the same step unless the user explicitly requests the full cycle and evidence is sufficient.
Required quality
The candidate must contain:
- minimal context;
- direct observation;
- reusable lesson;
- procedure;
- limits;
- confidence;
- evidence level;
- retrieval terms;
- source actor/session.
Never include private chain-of-thought, secrets, credentials, personal data, or raw conversation history.
Outcome
If no lesson qualifies, state NO_REUSABLE_LEARNING and do not call capture.
