In one sentence
Agent memory is whatever an AI can still use from earlier in a conversation or from past interactions, and by default it is far less than people assume.
The default: it forgets
A plain AI model does not remember anything between separate requests. Each call starts fresh. Within a single conversation it can refer back to earlier messages, but only because those messages are being re-sent to it each time, inside its context window. Once they scroll out of that window, or the conversation ends, they are gone.
This surprises people, because the agent feels like it remembers. It does not. It is re-reading.
Three kinds of "memory"
- In-conversation context: the recent messages, re-sent each turn. Limited by the context window. Reliable while it lasts, then forgotten.
- Stored memory: facts you deliberately save somewhere (a database, a file, a profile) and feed back in on later runs. This is real persistence, but only for what you chose to store.
- What the model learned in training: general knowledge baked in when the model was built. It does not include anything about your business or your customers.
Why this matters for design
If a workflow depends on the agent "remembering" something important, that memory must live somewhere you control, not in the model's head. For example, if an agent should recall a customer's earlier request next week, you store that request and retrieve it, often using retrieval. You do not assume the agent kept it.
What this does not solve
Storing memory raises its own questions: what you keep, for how long, and whether it is safe to hold. More memory is not automatically better, and personal data you store is data you must protect. Decide what genuinely needs remembering, not everything that could be.
Good enough to move on when
- You know what the agent needs to recall between runs, and where that lives.
- Nothing important depends on the model remembering on its own.
- You have decided what not to store, as deliberately as what to store.
