persistent-memorylisted

# Persistent Memory ## Overview Give the agent durable, cross-session memory through a simple file-based system. Avoid context amnesia by persisting facts, preferences, and decisions to disk, retrieving them on demand, and periodically distilling raw daily logs into a compact long-term record. ## When to Use - The user says "记住 X", "记下来", "长期记忆", "别忘记", "recall X", "remember", "memory", "跨会话上下文". - After completing substantive work that should survive future sessions. - When the user references something discussed in a prior session. ## Memory Layout Use a memory directory (default `./memory/` or the project's `.novera-ai-skills/memory/`). Two file types: - `YYYY-MM-DD.md` — append-only daily log for ephemeral, dated events. - `MEMORY.md` — curated long-term notes: durable facts, user preferences, project conventions, decisions. ## Procedures ### 1. Write (remember) - Durable fact / preference / convention → append a concise bullet to `MEMORY.md` (group by topic). - Ephemeral / dated event → append to today's `YYYY-MM-DD.md` (create with a date header if missing). - Keep entries atomic and scannable. Never store secrets unless explicitly requested. ### 2. Recall (search) - Identify topic keywords, then search the memory directory with Grep or `scripts/search_memory.py`. - Prefer `MEMORY.md` for durable facts; use daily logs for dated specifics. - Read the matched file(s) for full context before answering. ### 3. Distill (periodic) - When a daily log is older than 30 da