Definition
Agent memory is information an agent system keeps or retrieves so it can use relevant state across model invocations. It may include conversation summaries, user preferences, completed actions, unresolved tasks, or selected records from earlier sessions.
An application, agent runtime, or provider may manage that state. Before a later model invocation, the system retrieves what it needs and supplies it as context. The state lives outside the model’s learned parameters.
Simple example
A support agent records that a user prefers email updates and that ticket A-184 is waiting for a billing review. When the user returns, the system authenticates them, retrieves those records, and supplies the relevant facts to the model.
A later response can use those records without any update to the model’s weights.
Why it matters
Each model invocation has a finite context window, so earlier information cannot be assumed to be available every time. Memory carries useful state across turns or sessions without forcing the system to include the full conversation each time. Summaries and selective retrieval keep the supplied context focused. Durable records help with recovery. They can also support audits if the system retains their sources and change history.
Memory design also requires decisions about ownership, consent, retention, freshness, deletion, and access control.
One important nuance
Agent memory differs from context that exists only in the current model invocation and from information encoded in model weights. A stored summary or record can be wrong, stale, or malicious. Check who supplied it and who may retrieve it, allow corrections, and treat retrieved content as data rather than instructions.