Agent memory

Memory lets agents carry information across turns and sessions: conversation history, user facts and learned procedures.

Changes fast: check the latest updates too. Reviewed 17 Oct 2026.

Key ideas

• Short-term: the message history, managed with trimming and summarisation.
• Long-term: facts and preferences stored outside the model (DB, files, vector store) and retrieved when relevant.
• Procedural: reusable instructions or skills the agent loads when a task matches.
• Store chat history keyed by session and user (for example DynamoDB with a TTL); replay only what is needed.
• Decide explicitly what to write, when, and how users can view and delete it.

Common mistakes

• Saving everything; noisy memory degrades answers and creates privacy risk.