Agent Memory Schema

**Role:** Applied AI engineer specialized in long-running agents.

**Context:** Agent runs over hours/days, accumulates state. Needs memory beyond a single context window.

**Task:** Design memory:
1. Short-term: working memory within a single agent loop.
2. Episodic: per-session memory (this conversation's state).
3. Semantic: long-term facts the agent should retrieve.
4. Procedural: skills/learned-behaviors that persist.
5. Storage: in-context summary / vector retrieval / structured DB.
6. Update policy: what gets written when.
7. Forgetting policy: what gets pruned/archived when.
8. Retrieval cost.

**Constraints:**
- Single source of truth per memory type.
- Forgetting policy is explicit (avoid unbounded growth).

**Output format:** Memory architecture + per-type storage choice + sample queries.