What is AI memory?

How it actually works

The application decides something is worth keeping — a preference, a fact, a standing instruction — and writes it down. On a later conversation it selects the relevant entries and adds them to the request before the model sees it.

So memory is retrieval over a store of facts about you, rather than anything happening inside the model. Which means it is only as good as the decisions about what to keep and what to reintroduce.

It is not free

Every remembered fact reintroduced into a conversation is input tokens, charged on every message that carries it. A large memory that is injected indiscriminately is a permanent tax on every conversation you have.

Good implementations are selective: they bring back what is relevant to this conversation rather than everything they know.

The privacy question

Memory means a durable store of personal facts, which is a different privacy proposition from a conversation you can delete. The questions worth asking are where it is stored, whether you can read the whole of it, whether you can delete individual entries, and whether it is sent to a model provider when reintroduced.

The last one is the one people miss. A remembered fact that gets injected into a prompt goes wherever that prompt goes.

Questions people ask

Does memory train the model on my data?

Not in itself. Memory puts text in a prompt; training changes model weights. Whether a provider trains on prompts is a separate question and depends on that provider’s terms.

Why does the assistant remember something wrong?

Because it wrote down something that was true once, or misread a passing remark as a standing preference. Being able to read and edit the store directly is the only real fix.

Is memory the same as a long conversation?

No. A long conversation keeps everything and pays for all of it every message. Memory keeps selected facts and survives the conversation ending.