메뉴
Best practices for writing and maintaining high-quality role memories.
기본적으로 소스를 먼저 확인하는 Prompt가 선택됩니다. 직접 명령으로 전환하거나 로컬 사본을 다운로드할 수도 있습니다.
설치 여부를 결정하기 전에 SKILL.md와 SkillsMP에 표시된 보조 파일을 읽어 보세요.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
직접 명령은 검토 Prompt를 거치지 않습니다. 실행하기 전에 소스를 확인하세요.
npx skills add https://github.com/Dwsy/agent --skill memory-best-practices명령은 한 줄로 유지됩니다. 복사하기 전에 가로로 스크롤해 전체 내용을 확인하세요.
로컬 사본을 원하시나요? SkillsMP에서 현재 제공할 수 있는 파일을 다운로드하세요.
Operate Glimpse-APPs with progressive context — app list in system, domain detail via tools on demand.
Load prior knowledge from role memory before starting any task.
Use when user asks to design, choose, explain, or implement UI transitions, motion patterns, easing/timing choices, state-change animations, shared-element effects, drill-down navigation, tab/filter/list/detail switching, or vague requests like "make this feel smoother", "more iOS", "more natural", or "animate this UI" that need concrete motion taxonomy and implementation guidance.
SKILL.md 표시 중
| name | memory-best-practices |
| description | Best practices for writing and maintaining high-quality role memories. |
| whenToUse | When writing or reviewing memory entries, or when the user asks about memory quality standards. |
A quality standard for writing high-quality memories that compound in value.
[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.
[Nx] Increments| Trigger | Effect |
|---|---|
memory({ action: "reinforce", content: "..." }) | Manual +1 |
memory({ action: "search", query: "..." }) score ≥0.5 | Auto-reinforce +1 |
| Pending auto-promote (search ≥0.5) | Moves to consolidated at [0x] |
Reinforce ≠ Promote:
[Nx] within consolidated (used+1)Before saving:
✅ "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
| 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.
Each learning has LLM-auto-extracted tags. You don't tag manually.
Tags are used in search:
Example: searching "安全删除" finds entries tagged filesystem, safety, delete even if they don't contain those words.
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.
| 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.
| 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.
[3x]+ entries[0x] count (low noise)[0x] never reinforced (noise)Add learning: memory({ action: "add_learning", content: "..." })
Add preference: memory({ action: "add_preference", content: "...", category: "..." })
Reinforce: memory({ action: "reinforce", content: "..." })
Search: memory({ action: "search", query: "..." })
List: memory({ action: "list" })
Consolidate: memory({ action: "consolidate" })
LLM tidy: memory({ action: "llm_tidy" })
SOC 직업 분류 기준
