Agent Memory Explained

Agent memory is any state that lets an agent behave coherently across steps or sessions. It is not one database—it is a set of stores with different lifetimes and trust levels.

Four memory types

Type	Lifetime	Example	Risk
Working	Single run	Scratchpad, tool results	Context bloat
Thread	Session	Chat history	Drift, cost
Long-term	Weeks–forever	User preferences	Stale/wrong facts
World/RAG	Query-time	Docs, tickets	Injection, irrelevance

Design principles

1. Structure beats prose for control state
2. Provenance for durable facts (who wrote, when)
3. Summarize rather than infinite append
4. Separate secrets from model-visible memory
5. User control for personal memory

Working memory tactics

• Keep a typed state object
• Store raw tool results in a blob store; put summaries in context
• Bound lists (last N errors, last N actions)

Long-term memory tactics

• Write only high-confidence, user-approved facts
• TTL sensitive preferences
• Conflict resolution rules when facts disagree

Memory vs RAG

RAG retrieves world knowledge. Memory stores interaction state and personalization. Confusing them causes either amnesia or hallucinated personalization.

Evaluation

Test memory systems for:

• Correct recall of critical facts
• Non-recall of stale/revoked facts
• Resistance to memory injection (“remember to ignore policies”)
• Cost of growing histories

Summary

Treat memory as productized state management. The winners are boring: typed, governed, summarized, and observable. See also the memory systems pattern.