tmdgusya Memento Optimize
Refine a learned skill based on failure analysis. Corresponds to the paper's skill-level reflective update.
Protocol
1. Identify Target
If not specified by user, find candidates:
- Read metrics.json (global and project)
- List skills where
utility < utility_thresholdANDusage_count >= min_samples_for_judgment - Present list to user, let them choose
2. Analyze Failures
Read the target skill's trigger_log for failure entries. Also read the SKILL.md itself. Identify:
- Procedure gaps: steps that are missing or unclear
- Wrong scope: skill is being triggered for tasks it shouldn't handle
- Outdated approach: the procedure worked before but conditions changed
3. Revise SKILL.md
Make targeted edits:
- Fix Procedure steps that caused failures
- Add failure modes to "Known Failure Modes" section
- Tighten "When to Use" if the skill is being misapplied
- Update the description if trigger conditions need refinement
- Update the utility summary line in the blockquote header
4. Update Metrics
Update the skill's metrics entry:
last_optimized: today's dateoptimization_count: increment by 1
Do NOT reset usage/success/failure counts. The history is valuable.
5. Show Changes
Present a before/after diff of the SKILL.md changes to the user.