Agent Memory Ops

Use this skill when you need to keep an agent's memory layer healthy instead of letting MEMORY.md rot.

What it does

• scans MEMORY.md + memory/*.md
• detects duplicate / near-duplicate bullets
• extracts memory candidates from recent daily notes
• surfaces active follow-ups / TODOs
• filters obvious secrets from suggested memory output
• produces a concise maintenance report you can act on

Good use cases

• "帮我整理 MEMORY.md"
• "检查记忆层有没有重复和过期信息"
• "把最近几天的重要内容沉淀到长期记忆"
• "做一次 agent memory audit"
• "维护长期记忆 / daily memory / notebook memory"

Commands

Run from the workspace root that contains MEMORY.md and memory/.

python3 {baseDir}/scripts/memory_ops.py report --root .
python3 {baseDir}/scripts/memory_ops.py dedupe --root .
python3 {baseDir}/scripts/memory_ops.py digest --root . --days 7
python3 {baseDir}/scripts/memory_ops.py digest --root . --files 5 --format json

Recommended workflow

1. Run report to see gaps, duplicates, and follow-ups.
2. Run digest on the last 5-7 daily notes.
3. Promote only durable facts into MEMORY.md.
4. Keep volatile chatter in daily notes.
5. Never copy secrets into curated memory unless the user explicitly asks.

Output policy

• Prefer --format markdown for human review.
• Prefer --format json when another tool or script will consume the result.
• The script intentionally redacts / skips likely secrets from digest suggestions.

References

• references/playbook.md