THE DIRECT ANSWER

Agent memory is information retained or retrieved across steps or sessions to support later behaviour. It may include explicit records, conversation summaries or external storage. Define what is stored, how it is corrected and when it expires instead of treating memory as an unlimited human-like faculty.

Separate three concepts

Conversation history records previous exchanges. Context is the material supplied for the current model request. Memory is selected information stored and retrieved across interactions. These can overlap, but saving a fact does not guarantee it will be present or correctly used later.

Store deliberately

A preferred report format may remain useful for months. An uncertain inference or temporary deadline may not. Record the origin of important memories and give users a way to correct them. Avoid treating a casual remark as a permanent fact about a person.

Test the lifecycle

Ask an assistant to remember a naming preference, apply it later, change it and then delete it. Verify each stage. In a shared product, separate personal preferences from organisational knowledge and maintain access boundaries. A useful memory system reduces repeated explanation without making decisions mysterious. It should be possible to identify the stored information behind an action and stop using it when it is no longer relevant or valid.

A worked example

A project assistant stores the user's preferred report format. Later the user changes that preference. Test whether the new instruction replaces the old one and whether the system can explain which preference it is using. Keep temporary task notes separate from durable records. Retaining more information is not automatically useful when outdated or inappropriate details influence future actions.

COMMON QUESTIONS

Questions readers ask

Is an agent's memory always accurate?

No. Stored summaries can omit or distort information, and old facts can become stale. Provide correction, deletion and clear boundaries around what is retained.

SOURCES & NOTES

Follow the evidence.

AI-assisted explanation with linked technical references. Examples are illustrative learning exercises, not reported customer deployments or experimental results.

  1. anthropic.com — official reference

Public sources support the factual background. Analysis and illustrative exercises are identified in the text. How we assess AI · Editorial standards

Spot a factual error or an outdated detail?