What is an AI chat app with memory?

Most chat interfaces are stateless by default. A conversation lives in one thread, and the next thread starts blank. That is fine for a one-off question and exhausting for real work, because the context that makes an answer good — the constraint you mentioned last week, the audience you are writing for, the thing you already tried — has to be rebuilt every time.

What does memory actually store?

A useful memory system stores two different things. The first is durable facts: your role, the tools you use, names you refer to without explaining, preferences about tone or format. The second is working context: what a project is, what has been decided, what is still open. Facts change rarely; working context changes constantly.

Both are different from chat history. History is a transcript you can search. Memory is a distillation that is already in play before you type anything. A product that only offers history makes you do the retrieval yourself.

What should you be able to do with it?

• Read it. If you cannot see what the app believes about you, you cannot trust its answers.
• Correct it. Memory that is wrong is worse than no memory, because it is applied confidently.
• Delete a single entry, not just everything. Selective forgetting is the difference between a tool and a liability.
• Turn it off. Some conversations should not persist, and that should be one switch, not a support request.

Where memory goes wrong

The failure mode is not forgetting. It is remembering the wrong thing at the wrong moment — dragging an old preference into a new project, or treating a throwaway remark as a standing instruction. Good memory is selective. It retrieves on relevance to the current turn, not on recency alone, and it does not treat everything it has ever seen as equally true.

The second failure is invisibility. If a system quietly accumulates inferences about you and never shows them, you get the downsides of being known without the ability to correct the record. Any memory feature worth using has a settings page behind it.

How Swirl handles it

Swirl keeps a memory store per account and pulls the relevant parts into a conversation as it runs. The Memory tab lists what is stored, lets you add an entry by hand, edit or delete one, and switch the whole thing off. Memory is a paid-plan feature; the free plan is there to try the chat itself.

The practical test is simple: start a new chat tomorrow, mention nothing, and ask for something that depends on what you said today. If the answer is right, memory is working. If you have to explain yourself again, it is not.