7 · Memory & State Management

Memory gives agents context; state management makes them consistent, personalized and reliable. This diagram lays out the memory layers an agent needs and where each is stored.

How it works

1. 1Short-term (working) memory: the current conversation, held in the context window.
2. 2Long-term (persistent) memory: stored outside the context, retrieved when needed via embeddings.
3. 3User memory: profile and preferences, enabling personalization.
4. 4Semantic memory: domain knowledge, documents, FAQs and policies.
5. 5LangGraph injects the right memories into prompts each turn.

Key points

• Different memories are stored in different systems (key-value, relational, vector, cache).
• Best practices: store only what's necessary, separate user vs app data, use TTLs and encryption.
• Memory turns a one-shot model into a coherent, personalized agent.
• AgentCore Memory provides this at enterprise scale.

Why it matters

Without memory, agents forget everything between steps. Layered memory is what makes them consistent, personalized and reliable over long tasks and sessions.

Frequently asked questions

What are the memory types?

Short-term (working), long-term (persistent), user (profile/preferences) and semantic (knowledge).

Where is long-term memory stored?

Outside the context window, usually in a vector database, and retrieved when relevant.