Dwsy Memory Best Practices
A quality standard for writing high-quality memories that compound in value.
The [Nx] Priority System
[0x] New → unverified, just added
[1-2x] Normal → confirmed 1-2 times
[3x]+ High Priority → battle-tested, frequently used
[Nx] = usage count. The best quality signal. [6x] survived 6 sessions — probably right. [0x] might be noise.
How [Nx] Increments
| Trigger | Effect |
|---|---|
role_exec({ op: "reinforce", args: { content: "..." } }) |
Manual +1 |
role_search({ query: "..." }) score ≥0.5 |
Auto-reinforce +1 |
| Pending auto-promote (search ≥0.5) | Moves to consolidated at [0x] |
Reinforce ≠ Promote:
- Reinforce = increment
[Nx]within consolidated (used+1) - Promote = move entry from pending.md → consolidated.md (at [0x])
Learning Quality Checklist
Before saving:
- Durable — true next month?
- Non-obvious — changes approach next time?
- Actionable — what to do differently?
- Concise — under 120 chars?
- Not redundant — search first, reinforce if similar
Good vs Bad
✅ "MyBatis-Plus getOne needs .last('LIMIT 1') to avoid TooManyResultsException"
✅ "禁止 rm，优先 trash"
✅ "ACP: Agent 无状态，Client 持状态"
❌ "用户让我修了一个 bug" ← too generic
❌ "Error at /src/index.ts:42" ← copy-paste, no insight
❌ "服务器现在在 3000 端口" ← temporary state
Preference Categories
| Category | When |
|---|---|
Communication |
Language, style, tone |
Code |
Style, conventions, abstraction |
Tools |
CLI, editors, workflows |
Workflow |
Process, review, deployment |
General |
Everything else |
Rules: one per line, be specific, correct category.
Tags
Each learning has LLM-auto-extracted tags. You don't tag manually.
Tags are used in search:
- Exact tag match → +0.3 score
- Related tag (association graph) → +0.15 score
- Conceptually related entries surface even with different wording
Example: searching "安全删除" finds entries tagged filesystem, safety, delete even if they don't contain those words.
Two Paths to Memory
Path 1: Auto-extract (agent_end / compaction)
→ pending.md [○] → search ≥0.5 → promote → consolidated [0x]
→ reinforce over time → [3x]+ High Priority
Path 2: Manual (memory tool)
→ consolidated [0x] directly (you verified it's worth keeping)
→ reinforce over time → same priority ladder
Why pending? Auto-extracted memories are noisy. Pending layer filters by actual usage.
Consolidate vs LLM Tidy
| Operation | Method | What it does |
|---|---|---|
consolidate |
Rule-based | Exact + Jaccard dedup. Safe, never deletes unique entries. |
llm_tidy |
LLM-guided | Rewrites verbose, detects contradictions, suggests deletions. |
Use consolidate for routine maintenance. Use llm_tidy for deep cleanup.
Knowledge vs Memory
| Dimension | Memory (consolidated.md) | Knowledge (knowledge/) |
|---|---|---|
| Scope | Cross-session insights | Reusable patterns, decisions |
| Format | One-line entries | Full markdown files |
| Example | "禁止 rm，优先 trash" | Full design pattern with code |
| Tool | memory |
knowledge |
| Share | Per-role | Role / global / project |
Write knowledge when: full pattern with code examples, or share across roles.
Signs of Healthy Memory
- High ratio of
[3x]+entries - Low
[0x]count (low noise) - Entries under 120 chars
- Well-categorized preferences
- No duplicates or contradictions
Signs of Unhealthy Memory
- Many
[0x]never reinforced (noise) - Very long entries
- Duplicates saying same thing differently
- Contradictory preferences
- Outdated information
Quick Reference
Add learning: role_exec({ op: "add_learning", args: { content: "..." } })
Add preference: role_exec({ op: "add_preference", args: { content: "...", category: "..." } })
Reinforce: role_exec({ op: "reinforce", args: { content: "..." } })
Search: role_search({ query: "..." })
List: role_exec({ op: "list" })
Consolidate: role_exec({ op: "consolidate" })
LLM tidy: role_exec({ op: "llm_tidy" })