memory

Context & pattern retention for long-running work. Stores decisions, repeated failure patterns, conventions, open threads. Sharper over time, not larger. Activates on: remember, context, history, decision log, what did we decide, recurring issue, lesson learned, knowledge.

Memory — Context & Pattern Retention

에이전트와 일할 때 가장 짜증나는 부분 — 같은 대화를 두 번 하는 것. 이 스킬은 그 비용을 줄인다.

보존 대상:

  • 중요한 프로젝트 결정
  • 반복되는 실패 패턴
  • 유용한 가정
  • 미해결 리스크
  • 다음 세션으로 가져갈 가치 있는 컨텍스트

모든 것을 기억하는 게 아니라, 재발견 비용이 큰 것 만 기억한다.

Principle 0 — Save what reduces future friction

Do not store everything. Store what will matter again.

Good candidates:

  • architecture decisions
  • stack conventions
  • deploy constraints
  • naming rules
  • repeated bugs or failure patterns
  • user preferences that affect future execution
  • things that caused avoidable re-explanation

Bad candidates:

  • trivial one-off details
  • noisy intermediate states
  • temporary scraps with no future value

What to remember

Four types of memories that matter:

  1. Decisions — framework, deploy platform, auth provider, migration policy, etc. Should capture: what was decided, why, what tradeoff was accepted, when to revisit.

  2. Repeated failure patterns — build breaks, deploy failures, package conflicts, framework incompatibilities. High-value because they reduce wasted loops.

  3. User preferences — UI style, communication tone, approval requirements, fact-marking. These shape future agent behavior.

  4. Open threads — known gaps not yet prioritized, deferred risks, intentional postponements. Prevents forgotten debt from turning into repeated rediscovery.

Memory layers

Layer 1 — session notes (short-lived): what changed today, what broke, assumptions made, what needs attention next.

Layer 2 — durable project memory (longer-lived): stable decisions, repeated patterns, working conventions, high-value lessons.

Layer 3 — compressed knowledge (permanent): if something repeats 3+ times, condense it into a short reusable rule instead of keeping noisy logs.

Quality Log (per-artifact)

에이전트가 만든 결과물을 기록한다. 다음 세션에서 품질 추세를 보는 용도.

한 엔트리 = 한 결과물. 기록할 것:

  • Artifact: 파일 경로 또는 세션 ID
  • Type: code | doc | deploy | review
  • Outcome: shipped | reverted | blocked | needs-rework
  • Quality signal: 사람 피드백 (accept / reject / silent) + 자동 시그널 (tests pass/fail, build ok/error)
  • Lesson: 이 결과물에서 남길 한 줄

한 주에 3회 이상 같은 lesson이 반복되면 Layer 3으로 승격한다. 반복되는 품질 구멍이 durable 규칙으로 굳는 메커니즘.

Error Pattern Ledger (failure-specific)

반복되는 에러는 별도 원장에 모은다. Quality Log와 달리 실패 재현 가능성에 초점.

한 엔트리 = 한 에러 패턴. 기록할 것:

  • Error signature: 에러 메시지의 스테이블한 부분 (정규식 가능)
  • First seen / Last seen: 날짜
  • Occurrence count: 누적 발생 횟수
  • Root cause category: code | env | schema | deploy | external | unknown
  • Fix: 가장 짧은 재현 가능한 수정
  • Prevention: 재발 방지책 (테스트 추가 / 린트 규칙 / 체크리스트 항목)

Promotion 규칙: 같은 signature가 3회 이상 발생하면 build/ship의 Quick-Fix Catalog로 승격한다. 해결책이 카탈로그에 들어가면 Error Pattern Ledger에서는 "resolved: see build/ship"로 대체.

Memory record format

Fields: topic, type (decision / pattern / preference / open-thread), summary, why it matters, trigger, when to revisit.

Full examples → references/record-format.md

Storage and retrieval

Organized across three layers: CONTEXT_LOG.md (session decisions), LOGS/ (daily snapshots), and memory/knowledge/ (durable rules).

Full details → references/storage-structure.md

Compression rule

If the same thing comes up multiple times:

  • stop storing raw repetition
  • compress it into a general rule
  • keep the shortest useful version

Memory should get sharper over time, not just larger.

Retrieval rule

Before starting a related task, check whether relevant memory exists.

Especially for:

  • deployment work
  • auth changes
  • environment setup
  • repeated bug classes
  • product or UX preferences
  • strategic decisions already debated

The point is to reduce repeated questioning and repeated mistakes.

Anti-patterns

❌ storing everything
❌ storing vague summaries with no future use
❌ keeping raw noise instead of compressing lessons
❌ treating temporary confusion as durable memory
❌ remembering facts but not the reason behind them
❌ re-asking the user something that was already settled clearly

Output expectations

When this skill is applied, the result should help answer:

  • what should not be forgotten
  • what should affect future behavior
  • what should be reused automatically next time
  • what is still unresolved but worth keeping visible

This skill should make the next session lighter, not just longer.

Execution Patterns

상세 워크플로우 → references/execution-guide.md

다루는 시나리오:

  • 세션 결정사항 캡처
  • 에러 패턴 기록
  • 주간 변경 요약
  • 보존 vs 폐기 기준
  • Memory hygiene + cleanup

cowork-engine 와의 연동

solo-cto-agent knowledge <project-dir> 명령은 이 스킬의 포맷을 그대로 사용한다. 저장 위치: ~/.claude/skills/solo-cto-agent/knowledge/

추출된 ERROR_PATTERNS 는 failure-catalog.json 으로 자동 머지되어 다음 review 호출에서 즉시 활용된다.

공통 스펙 참조

  • 출력 포맷·팩트 태깅: skills/_shared/agent-spec.md
  • 임베드 컨텍스트: skills/_shared/skill-context.md

CLI Hooks — personalization 명시 기록

solo-cto-agent feedback accept --location src/Btn.tsx:42 --severity BLOCKER
solo-cto-agent feedback reject --location src/Nav.tsx:12 --severity SUGGESTION \
  --note "이미 memoized — false positive"
solo-cto-agent feedback show                                 # 누적 accept/reject 패턴
solo-cto-agent knowledge                                     # 결정/에러 패턴 추출
solo-cto-agent session save                                  # 세션 스냅샷

feedback 은 다음 review 의 가중치로 반영 (80/20 anti-bias rotation — 과거 패턴 과적합 방지). 상세: docs/feedback-guide.md.