Pattern · intermediate
Memory Systems Pattern
Patterns for agent memory: working state, summaries, long-term stores, and retrieval with governance.
Memory systems give agents continuity. The pattern is to split stores by lifetime and trust, not dump everything into the prompt.
Pattern blueprint
- State object — JSON schema for the run
- Episode log — trajectory events
- Summarizer — compress old steps
- Long-term KV / vector store — durable facts
- Retriever — fetch relevant world docs
Write policies
- What can be written automatically?
- What needs user approval?
- What is never stored?
Read policies
- What is injected every turn vs on demand?
- How many items max?
- How do you handle conflicts?
Anti-patterns
- Infinite chat history as the only memory
- Storing secrets in model-visible memory
- Unbounded “remember this” from untrusted text
Deep dive: Agent memory explained.
FAQ
Frequently asked questions
How should agents store memory?
Use typed working state for control flow, summarized thread history for context, and governed long-term stores with provenance for durable facts—plus RAG for world knowledge.