What Each AI Assistant Remembers, and for How Long

The usual version of this question is which AI has the best memory. It is the wrong question, because the six main assistants are not doing the same job well or badly. They are doing different jobs.

Here is what each one actually keeps, whether you have to switch it on, and what you can see or correct. Everything below comes from the companies' own documentation.

The short answer

ChatGPT keeps the most, and shows you the least of what it keeps. Claude gives you the most control, item by item. Gemini knows the most about you already, because it is sitting inside your Google account. Grok lists memory as a feature and behaves inconsistently in practice. DeepSeek keeps your chat history but starts every new conversation empty. Kimi keeps a small, capped set of notes you can edit in plain language.

None of them share anything with each other.

What each one keeps

ChatGPT. Memory is built from your conversations in the background rather than saved line by line, and OpenAI has said openly that what it shows you in the memory summary is not everything it remembers. The company has also acknowledged that older saved memories went stale over time and could contradict each other, the example being marathon training stored next to a sprained ankle. You can read the summary, switch memory off, or use a temporary chat that neither uses nor creates memories. Project only memory is a setting you turn on inside a project, and chats in that project then ignore your saved memories and cannot reference conversations outside it.

Claude. Memory is saved as individual topics while you are talking, not summarised afterwards, and you can open settings and edit or delete them one at a time. It is on by default on personal plans, and off by default for team and enterprise accounts until an administrator enables it. Each project keeps its own memory space, separate from other projects and from chats outside projects, which is the cleanest separation any of them offer. Searching your past conversations is a paid feature.

Gemini. Past chats are used by default for personal accounts, for users over eighteen, and the reach is wider than the chat window, because it draws on the Google account it lives in. It also offers a tool to import memories and chat history from other AI apps. Worth knowing: Google says conversations that have been seen by human reviewers are kept for up to three years even if you delete your activity, disconnected from your account. That distinction, and the equivalent policies elsewhere, is covered in where those conversations live.

Grok. The product page lists memory across chats as a feature. In our own testing on a free account with the setting on, facts stated in one conversation did not appear in a new one across three attempts. History syncs across web, iOS and Android, and private chats stay out of history and are removed within thirty days. Treat it as listed but inconsistent rather than present or absent.

DeepSeek. There is no official documentation describing cross chat memory, and in practice a new conversation starts blank. Chat history syncs across devices, and the context window inside a single conversation is now very large by default, so long threads hold together well. The strength is depth within one conversation, not continuity between them.

Kimi. A memory space that records preferences, habits and style, managed in plain language: you tell it to remember, update or forget something, or edit entries in settings. It is explicitly capped, around fifty entries of five hundred characters each, and it declines to store things like health details, passwords and addresses unless you insist. Small and legible rather than comprehensive.

What none of them do

Three gaps run across the whole list.

None of them share with each other. What you tell one, the others cannot see, and nobody is building that bridge. Gemini will import from other apps and Claude will import and export, which tells you how the industry sees this: bringing memory in is a feature, keeping it in sync going out is not. An import is a one time copy that starts drifting apart the next day. The reasons behind that, and what follows from it, are in what AI actually remembers.

None of them keep your words. Every one of these stores a condensed version: a summary, a topic, a preference. A summary knows you decided something. It rarely keeps the paragraph where you worked out why, and that paragraph is what you want three weeks later.

None of them are fully visible. Two of the six say outright that what you can see is not the whole of what they hold. Deleting a conversation does not reliably delete what was learned from it, because the memory is stored apart from the chat that produced it.

If you are choosing just one

If most of your work is one continuous project, Claude's per project memory keeps other work out of the way, and the item level editing means you can correct it when it gets something wrong.

If you want the assistant to accumulate a picture of you without being managed, ChatGPT does that most aggressively. The trade is visibility.

If your working life already runs on Google, Gemini starts with context nothing else has, and you should read its retention policy before you decide that is a good thing.

If you mainly need long single conversations rather than continuity, DeepSeek's large context holds a long thread together well.

If you want a memory you can read in one screen, Kimi's cap is a feature, not a limitation.

What this list does not tell you

It does not tell you what to do if the answer is more than one, which for most people it is by now. Drafting happens in one, code in another, research in a third, and the memory in each is invisible to the other two. No item on this list improves that, and no future update is likely to, because memory is one of the strongest reasons to stay with a single assistant.

That case is a different problem with different solutions: moving context by hand when it happens occasionally, or something that sits outside all of them when it happens every week. Moving context by hand covers the first, and which form solves which problem covers the second.

One more caveat. Every line above was checked against official documentation this month, and this is an area where the companies ship changes quarterly. Check the settings page in your own account before you rely on any of it.