What cross-AI memory actually means.

Every AI assistant available today keeps its own memory, if it keeps one at all, and that memory is maintained by that one assistant. Your phone and your laptop see the same history, because they are the same account. A different assistant does not. Claude offers a manual import of what another one had learned about you, which is a real convenience and is not the same as two assistants keeping one thing current between them. Say something useful to one, and it stays that one’s to remember. The manual fix people reach for first is obvious: keep a document with the important parts, and carry it with you: paste it in, attach it, whatever the assistant of the moment allows. It works, for a while, for anything short-lived.

A bigger memory inside one assistant. The most common reading, and not cross-AI at all. A longer context window, a better memory list, more files in one project: real improvements, every one of them owned by the assistant that made them.

Moving memory from one to another. Closer, and it exists. Claude takes a paste of what another provider had learned about you: preferences, your role, the things it noticed, and goals and projects too, since the export it asks for includes them. Anthropic calls the feature experimental and says Claude may not always incorporate what you paste. What it does not do is keep the two in step. It is a copy made by hand, and from the moment it lands the two carry on separately.

Exporting your chat history. Also real, and it is a transcript. It tells you what was said, in order, including everything you later changed your mind about. Reading it back is the work you were trying to avoid.

A document you carry between them. The manual version, and it works until it does not. It knows only what you remembered to put in it, as of the last time you updated it. Something changes, and whichever assistant you open next answers from a version that has quietly gone stale. It will not tell you it is stale. It will answer as if it is current, because as far as it knows, it is.

Something narrower than all four: a record that belongs to you rather than to any assistant, that more than one of them can read and write to, and that stays current as the thing it describes changes. Not a bigger memory, not a copy, not a transcript, not a file you maintain. Which assistant you happen to be talking to stops being the thing that decides what it knows.

Giraffe is one way to do that and not the only conceivable one. What matters for the definition is the shape: outside the assistants, readable by them, and kept current on purpose.

Making the same facts reachable everywhere is the easy half. What actually matters is what happens when something changes, and it will, because the whole reason you’re using more than one AI is that your thinking keeps moving. Before it adds something new, whichever assistant you’re in checks what the project already says. A fact that’s changed updates in place, the old version kept as history. A decision that gets reversed stays visible instead of vanishing, marked reversed, with the reason attached. None of that depends on which specific assistant you’re using in the moment: it’s a property of the project, not a feature of any one AI.

Say you’re weighing a couple of job offers. On your phone, over lunch, using whichever assistant you’ve got open: “save the decision: going with the smaller company, the flexibility matters more than the base salary difference, at least for now.” A few days later, at your desk, after a second conversation with the smaller company changes the picture: “reverse that decision: going with the larger one instead, they matched the flexibility and the difference in comp is too big to leave on the table now.” The first decision doesn’t disappear. It stays exactly where it was, marked reversed, with the reason attached, not quietly deleted the moment you changed your mind. That evening, back on your phone, you ask whichever assistant’s open to check the project before you draft a reply to the smaller company. It reads the larger-company decision as current, not because your phone remembers what you said at your desk, but because you told it exactly where to look.

Open the project on its own and it isn’t a log of which AI said what, when: it’s just the decision, in plain language, with the reasoning behind it. That’s the Simple view. Structured view shows the same project broken into its labeled parts. Neither one cares which assistant you happened to be using when any of it was said.

The longer you do this, the less it’s about any single project and the more it’s about all of them at once: a few hundred conclusions, across everything you’ve ever asked an AI to help you think through. Cross-AI memory that only solves “today, on whichever assistant” isn’t enough by itself. It has to still work when you’re trying to find the one decision you made eight months ago, in some AI you don’t even use anymore, about a job offer you can barely remember the details of. Search across every project you have, and it isn’t hunting through old conversations hoping you’ll recognize a sentence. It’s looking at what you actually kept: the facts, the decisions, the questions, wherever they came from.

Most people asking this question run two assistants, not five. Keeping one thing current across ChatGPT and Claude is the practical version of this page: what actually transfers between them, what does not, and what to do about the part that does not.

If you use Gemini as well, it connects through Spark rather than an ordinary chat. The setup guides cover each assistant and are kept current as their own settings change.