Agent memory is the information an AI agent keeps and reuses beyond a single model call: the steps already taken in a task, notes from earlier sessions, facts about a customer it has handled before. It lets an agent pick up where it left off. It also creates a store of data that needs the same governance as any other record.
What kinds of agent memory are there?
Three, broadly. Working memory is what sits in the model's context during one task: the instructions, the conversation so far, the tool results. It disappears when the task ends. Session memory carries a task across interruptions, such as a case that pauses for a customer's reply. Long-term memory is anything the agent writes down to use in future tasks: a summary of a customer's preferences, a note that a supplier always invoices late.
Where several agents read and write the same store, that is shared memory, and keeping it consistent becomes its own problem. See how to keep agents' shared memory consistent.
Why does memory matter once agents act?
Because a remembered note can become a reason for an action. If an agent noted last month that "the customer agreed to monthly statements by post", it may act on that note today. Was the note right? Who said it? Is it still true? Memory without provenance turns a passing remark into a standing instruction.
Memory is also personal data in most regulated settings. A note about a customer's circumstances is subject to the same rules as the customer file: lawful basis, storage limits, the right to erasure.
How should agent memory work in practice?
Treat each memory entry as a record, not a scratchpad. It should carry where the underlying fact came from and its provenance class, when it was written, and by which agent. A note the agent wrote summarising what a customer said is, at best, declared. It should never outrank the signed record it summarises.
Retention should be set per type of memory and enforced, not left to grow. Erasure requests should reach memory stores as well as primary systems. The details are covered in GDPR erasure requests and AI agent memory and GDPR storage limitation and AI agent memory.
What does that look like in an example?
A wealth manager's servicing agent remembers that a client prefers calls in the afternoon and once asked about ethical funds. Both notes are useful and low-risk; they shape drafts. The agent also remembers that the client "said they are happy with higher risk". That note could influence a suitability conversation, so it is marked declared, linked to the call it came from, and cannot change the client's recorded risk profile. Only an adviser can do that, through the signed suitability process.
How is agent memory different from a knowledge base?
A knowledge base is curated material the business has approved for agents to use, such as product guides and procedures, owned and maintained by people. Agent memory is what the agent itself recorded while working. The first is a source the agent is given; the second is the agent's own notes. Mixing them, for instance by letting agents write freely into the knowledge base, lets an agent's guess become something every agent treats as reference.
Questions people ask
Should an agent remember everything?
No. Keep what helps future tasks and has a clear purpose, and set a retention period for each kind. Anything that cannot be justified under the business's data rules should not be stored.
Can a person correct an agent's memory?
They should be able to, and the correction should be recorded. Memory that cannot be inspected or corrected is a liability, because nobody can tell why the agent acts the way it does. The governed agentic operating system guide covers how records are kept reviewable.