memento-reflect

Memento Reflect

Post-task self-evaluation that closes the learning loop. This is the Write phase of Read-Write Reflective Learning.

Protocol

1. Trace Analysis

Review the current conversation and identify:

• What task was performed?
• Which tools and approaches were used?
• Were any learned skills referenced or applied?
• How many iterations/retries were needed?

2. Self-Evaluation

Judge the outcome using these criteria:

• User accepted result without corrections? — strong success signal
• Completed in one pass? — success. Multiple retries — partial or failure.
• Errors encountered during execution? — note which phase failed
• Final result meets the original request? — ultimate success criterion

Verdict: success or failure with one-sentence reasoning.

3. Update Metrics

If a learned skill was used during this task, update its metrics file.

Read the appropriate metrics.json (~/.claude/memento/metrics.json for global skills, .claude/memento/metrics.json for project skills). Then update the skill entry:

• usage_count: increment by 1
• success_count: increment if success
• failure_count: increment if failure
• last_used: today's date
• utility: success_count / (success_count + failure_count)
• trigger_log: append {"task": "short description", "result": "success|failure", "date": "today"} — keep max 5 most recent

Write the updated JSON back. Create the directory and file if they don't exist yet — initialize with:

{
  "version": 1,
  "config": {
    "utility_threshold": 0.4,
    "min_samples_for_judgment": 3,
    "prune_after_days_unused": 60
  },
  "skills": {}
}

4. Pattern Discovery

Ask yourself: "Did I do something reusable that no existing skill covers?"

Criteria for a new skill:

• The pattern would apply to future tasks (not one-off)
• It required non-obvious steps or domain knowledge
• An agent without this experience would likely struggle

If yes — tell the user what pattern you found and suggest: "Want me to capture this as a learned skill? I'll use /memento create."

5. Failure-Driven Optimization

If a learned skill was used but the task failed:

• Was the skill's Procedure incomplete or wrong?
• Was it applied to the wrong type of task (routing problem)?

If the skill itself is at fault — suggest: "The {skill-name} skill may need improvement. Want me to run /memento optimize {skill-name}?"

Important

• Always show the user what you're updating before writing
• Never silently modify metrics — announce changes
• One reflect per task — don't batch multiple tasks