Memory System Ops

Run this skill when the user asks to improve, audit, or maintain memory quality.

Strict Runbook (for weaker agents)

When this skill is triggered, execute in this order:

1. Read current state files:
   • USER.md
   • TASKS.md
   • memory/YYYY-MM-DD.md (today)
   • optional: agent-memory/common-info.md
2. Set retrieval scope before searching (when applicable):
   • user / project / session / date range
   • avoid global scans if a narrower scope is enough.
3. Retrieve with efficiency gates (must follow):
   • L1: today memory + task board (+ memory/keyword-frequency.md for keyword questions)
   • run a sufficiency check: if answer is already supported by evidence, stop here
   • L2: scan memory/*.md only when L1 is insufficient
   • run a sufficiency check again
   • L3: fallback to agent-memory/chat/daily -> weekly -> monthly only when confidence remains low
4. For every important new decision, append one structured block:
   • Decision / Why / Impact / Next / Verify
   • optional but recommended: RefID / Scope / Confidence / Tags / Aliases
5. Update TASKS.md with only 在做 / 就绪 / 中断.
   • If 中断, must fill Blocked by and Resume step.
6. After memory write-back, update memory/keyword-frequency.md.
7. Apply checkpoint policy (for /new-resilient mode):
   • Event trigger (mandatory): on new preference/boundary, task state change, or key external action completion, append write-back immediately.
   • Pre-refresh trigger (mandatory): before /new / /reset / context refresh, append a lightweight checkpoint first.
   • Periodic fallback (optional but recommended): every 30–45 minutes, append a lightweight checkpoint with a min-interval guard (e.g., skip if last checkpoint < 25 minutes).
   • If no new decisions, still write a minimal checkpoint (timestamp + TASKS snapshot).
8. If the same rule/decision repeats >=3 times, compact and promote:
   • long-term preference -> USER.md
   • operational rule -> agent-memory/common-info.md
   • keep a reference link to original daily entries.
9. Reply with verifiable evidence (file path + what changed), not vague claims.

Never skip step 3 for historical recall questions.

Core Model

Use 3 layers:

• USER.md: stable user profile and long-term preferences.
• memory/YYYY-MM-DD.md: daily facts and decisions.
• TASKS.md: active executable tasks only (在做/就绪/中断).

Use TASKS_ARCHIVE/YYYY-MM.md for completed tasks.

Daily Write-Back Standard

For key events, write:

• Decision
• Why
• Impact
• Next
• Verify
• optional: RefID, Scope, Confidence, Tags, Aliases

When the user reports memory mismatch ("你没记住"), immediately show the exact stored location and verify against file content.

Checkpoint Standard (frequent /new support)

Use this lightweight checkpoint block when there is no major new decision but state continuity is needed:

• RefID (e.g., AUTO-YYYYMMDD-XX)
• Decision (auto flush checkpoint)
• Why (protect continuity before reset/switch)
• Impact (recoverable context after reset)
• Next (continue normal write-back loop)
• Verify:
  • checkpoint time
  • TASKS.md snapshot counts (在做/就绪/中断)
  • optional top active task IDs

Goal: preserve recoverability without bloating daily notes.

Retrieval Fallback Order

1. L1: today memory + task board
2. L2: memory/*.md
3. L3: agent-memory/chat/daily -> weekly -> monthly (when low confidence)

Keyword-first rule (must apply)

For keyword questions, do this before final answer:

1. Read memory/keyword-frequency.md for candidate terms and counts.
2. Locate matching daily entries in memory/*.md.
3. Return concrete path evidence (where the keyword/decision appears).

Always prefer verified answers over guessed answers.

Efficiency Enhancements (memU-inspired)

1. Pipeline stages, not one-shot search
   • Run retrieval as staged execution (L1/L2/L3 + sufficiency checks).
2. Scope guard first
   • Narrow where scope before retrieval to reduce noise and latency.
3. Sufficiency gate
   • Stop retrieval once evidence is enough; avoid expensive unnecessary scans.
4. Reference-friendly write-back
   • Add short RefID in important entries so later answers can cite quickly.
5. Reinforcement + compaction
   • Repeated decisions should be promoted to stable layers to prevent daily-file bloat.

Task Board Rules

• Keep only active states in TASKS.md: 在做/就绪/中断.
• For 中断, require:
  • Blocked by
  • Resume step
• Move completed tasks to TASKS_ARCHIVE/YYYY-MM.md.

Rule Placement

• Keep user identity and stable preferences in USER.md.
• Keep operational rules (migration/archive/retrieval) in agent-memory/common-info.md.

Keyword Tracking

• Maintain keyword seeds manually in memory/keyword-seeds.md.
• Generate stats to memory/keyword-frequency.md after each daily write-back.
• Default mode is event-driven.
• For high-frequency /new workflows, you may enable optional periodic checkpoint mode; if enabled, keep a min-interval guard to avoid noisy duplicate entries.

References

• Workflow details: references/workflow.md
• Daily template: references/daily-template.md
• Task board template: references/tasks-template.md
• Efficiency notes: references/efficiency-upgrades.md
• Checkpoint mode: references/checkpoint-mode.md