feat: Agent memory — per-agent persistent knowledge across sessions

[JeremyDev87]

Purpose

에이전트별 누적 지식 저장. security-specialist가 이전 세션에서 발견한 패턴을 다음 리뷰에 자동 반영.

Changes

• Create: packages/claude-code-plugin/hooks/lib/agent_memory.py
• Create: packages/claude-code-plugin/tests/test_agent_memory.py
• Modify: packages/claude-code-plugin/hooks/stop.py — 세션 종료 시 agent memory 업데이트

Acceptance Criteria

• ~/.codingbuddy/agent_memory/ 디렉토리에 에이전트별 JSON 파일 저장
• 에이전트 이름 기반 키 (e.g. security-specialist.json)
• 저장 항목: findings (발견사항), patterns (반복 패턴), preferences (선호 방식)
• 최대 항목 수 제한 (FIFO, 기본 50개)
• Stop hook에서 세션의 agent 활동을 memory에 기록
• SessionStart에서 관련 agent memory를 system prompt에 주입 가능
• TDD: 최소 10 테스트

[m13v]

agent memory is one of those things that seems simple but the storage format matters a lot. we built a persistent agent knowledge system and initially used plain JSON files like you're proposing. it works fine until you need to query across agents or do any kind of semantic retrieval. the FIFO limit of 50 items is a good start but you'll want some kind of importance scoring too - not all findings are equally useful in future sessions. what worked for us: SQLite with full-text search for exact lookups plus embeddings for semantic similarity. agents store structured facts (not raw conversation logs) with confidence scores that decay over time. the stop hook approach for saving is right, but also consider saving incrementally during the session - if the agent crashes, you lose everything otherwise.

[m13v]

our agent memory implementation using SQLite + full-text search: https://github.com/m13v/ai-browser-profile/blob/main/ai_browser_profile/db.py

and the embeddings layer for semantic similarity retrieval across stored knowledge: https://github.com/m13v/ai-browser-profile/blob/main/ai_browser_profile/embeddings.py