Memory Isn’t State
Dylan Telano · Oct 23, 2026 · 2 min read
TL;DR: AI agent memory remembers everything, including what is no longer true. State keeps only what still applies. Agents need both, and most only have one.
AI agent memory vs state, in one example
On Monday, your CFO caps renewal discounts at 10%. The old policy said 15%. On Thursday, someone asks your sales agent whether they can offer a customer 15% to close.
An agent with good memory has both facts. It remembers the 15% policy, because it read it for months. It remembers the Monday email too. Which one wins depends on what gets retrieved, how the prompt is ordered and a little luck. Sometimes it says yes.
An agent with state does not have to guess. The 15% rule was superseded on Monday. The current rule is 10%. That is the only one it is handed.
Remembering is not knowing
Memory answers “what have I seen?” State answers “what is true now?” They sound close. They behave very differently as a task gets longer.
| Memory | State | |
|---|---|---|
| Keeps | Everything it has seen | What still applies |
| When a fact changes | Adds the new one next to the old | Replaces the old one, and records why |
| Over a long task | Grows, and gets noisier | Stays current |
| Failure mode | Retrieves the wrong version | Missing a fact it was never told |
| Best at | Recall, personalisation, history | Decisions, policies, progress, commitments |
Why more context does not fix it
The usual fix is more: a bigger context window, a better retriever, smarter context engineering. All of it helps the model see more. None of it tells the model which of the things it sees are still true. Put two contradictory policies in front of a model and you have not given it memory. You have given it a coin flip.
A good test you can run on your own agent: change one rule halfway through a long task, then ask a question that depends on it. Then do it again with three changes. Watch which version it uses.
You want both
This is not memory against state. A good agent needs to recall the conversation, the customer and the history. It also needs a layer that keeps decisions, rules and progress current, and marks what was replaced. Memory layers are getting good at the first job. Cricket does the second.
Memory is what you have seen. State is what is true. Your agents are being asked to act on the second.