📊 Mnemosyne Recall 品質監控（每週） LLM P2 risk=LOW

1. 做什麼 / 2. 為什麼存在

Mnemosyne Recall 品質監控（每週）：抽樣評估召回品質。

3. Schedule / Mode

mode=LLM state=scheduled enabled=1
schedule=20 7 * * 0
next_run=2026-08-30T07:20:00 last_run=2026-08-23T09:01:17
deliver=telegram model=gpt-5.6-luna provider=openai-codex

4. Implementation

script=—
monitor_script=— monitor_url=—
skills=[]
workdir=—

5. Effective instruction sources

prompt_sha256=dd84c02cd1105c57… chars=412
skill_files={}

執行 mnemosyne_recall_diagnostics(reset=true) 取得上週的 recall metrics。

從回傳數據中檢查以下異常信號（任一觸發即回報）：

1. **fallback rate > 10%**：fallback 調用佔比過高，表示語義搜尋經常找不到結果
2. **total 調用數異常下降**：與上週相比下降 > 50%（若無基準數據，跳過此項並標註「尚無基準」）
3. **任何 tier 的 hit count = 0 但上週 > 0**

若所有指標正常，回覆 "[SILENT]"。
若有異常，回覆簡短摘要（指標名 + 數值 + 建議行動）。

同時記錄本週的 total 調用數到 ~/.hermes/data/mnemosyne_recall_baseline.txt（格式：YYYY-MM-DD|total_calls|fallback_rate），供下週比較。…

6. Actual rules / thresholds

rule	value	source
none extracted

7/8. Inputs & Outputs (resources)

kind	path	name	access
none

9/10. Upstream / Downstream

Upstream: —
Downstream: —
Transitive downstream (0): —

Edges (typed evidence)

type	hard?	from	to	resource	origin	conf	evidence
none

11. Shared resources

12. Recent executions

failure_streak=0
last_error=—
delivery_error=—

14. Drift history

2026-08-23T09:28 JOB_ADDED job 5f490634fad5 appeared (📊 Mnemosyne Recall 品質監控（每週）)
2026-08-23T08:42 JOB_ADDED job 5f490634fad5 appeared (📊 Mnemosyne Recall 品質監控（每週）)
2026-08-23T08:41 JOB_ADDED job 5f490634fad5 appeared (📊 Mnemosyne Recall 品質監控（每週）)
2026-08-23T08:36 JOB_ADDED job 5f490634fad5 appeared (📊 Mnemosyne Recall 品質監控（每週）)
2026-08-23T08:17 JOB_ADDED job 5f490634fad5 appeared (📊 Mnemosyne Recall 品質監控（每週）)
2026-08-23T08:17 JOB_ADDED job 5f490634fad5 appeared (📊 Mnemosyne Recall 品質監控（每週）)

16. Impact / change risk

change_risk=LOW