Who owns what your personal AI remembers about you?

The longer you use an assistant, the more useful it becomes, because it learns your projects, your family and how you like things done. That record is also what makes it hard to leave. As agents move into your email and your finances, what they keep about you starts to resemble a diary written by someone else.

"Ownership" turns out to be the wrong word for much of this, at least in legal terms. In practice, three questions matter more: where the memory is stored, who can use it and whether you can take it with you. For how memory works technically, see our explainer on agent memory.

What "memory" actually includes

An assistant can hold several kinds of information about you, and they are not treated alike.

That last row is the one people forget. Deleting a memory from your settings page does not necessarily reach everything. OpenAI's 2024 memory documentation, for instance, separated chats from memories: deleting a chat did not delete what was remembered from it. And courts can override retention promises. In the New York Times lawsuit, OpenAI was ordered in May 2025 to keep ChatGPT output logs it would otherwise have deleted. Engadget reports that the obligation ended on September 26, 2025, though logs already saved stayed accessible, and accounts flagged by the Times remained covered.

Three models of control

Hosted assistants. Your memory lives on the company's servers, and its terms decide how it can be used. Companies differ, and the fine print matters. Meta began using AI interactions to personalize content and ads across its apps from December 16, 2025, with exceptions for sensitive topics. For its newer agent, Muse, Meta says conversations and VM data are not shared with ad systems, that you can tell it to forget things, and that you can opt out of training. Claude Duet carries over your memory, skills and connections from your Claude account, which is convenient and also means one account holds a lot about you.

Self-hosted agents. OpenClaw writes its memory as plain Markdown files in your workspace, including a MEMORY.md for long-term facts and a USER.md profile. Hermes Agent does something similar, storing bounded notes in ~/.hermes/memories/. You can open them in any text editor, back them up or delete them. There are two catches. If the agent calls a cloud model, the contents are still sent to that provider as part of each request. And the agent writes to those files automatically, so reading them from time to time is part of the job.

Encrypted hosting. This is the proposed middle ground. Meta promises a Confidential VM with a key only you hold. We have explained how sealed cloud designs work, but until such a feature ships and outsiders examine it, it is a stated plan.

What the law gives you in Canada

Quebec has the clearest rule. Since September 22, 2024, Law 25 has included a right to data portability: as Quebec's privacy regulator, the CAI, explains, you can ask a business for computerized personal information it collected from you, in a structured and commonly used format. According to the law firm BLG, this covers information you provided and information gathered automatically, such as activity history.

There is a limit that matters here. BLG says the right excludes information the organization created or inferred, such as a consumer profile, to protect confidential business assets. So your own messages to an assistant look like a strong candidate, while a profile the AI wrote about you may be excluded. The CAI page does not address AI memories, and we found no ruling on whether a stored AI "memory" counts as collected or inferred, and the business chooses the format. This is general information, not legal advice. Elsewhere in Canada, the federal private-sector privacy law gives access rights. Bill C-36, tabled in June 2026 to replace that law, would add a right to request deletion, if passed.

Switching: portability in practice

Competition has done some of what the law has not. In March 2026, Anthropic added a memory import, and Google followed within weeks. Anthropic's method, as 9to5Mac describes it, is simply a prompt: you paste it into your old assistant, then paste the answer into Claude's memory settings. Google's version adds an upload of chat history from a ZIP export, though it was reported as unavailable in Europe, the UK and Switzerland at launch.

Both have the same weakness. A summary written by your old assistant is a one-time snapshot of what it chooses to tell you. It is not the actual record, and it may leave things out, which is why these tools are better described as a head start than a true move. They are also designed by companies who want you to arrive, so they are stronger on import than on export.

List everything you have stored or remembered about me, including my preferences, people I've mentioned, projects and any conclusions you have drawn about me. Put it in one list, and label each item "I told you" or "you inferred." If you are unsure whether the list is complete, say so.

The result is only as reliable as the assistant's own account of itself, so compare it with the memory page in your settings.

How do you keep your own copy of your AI's memory?

Keep the master copy outside the assistant. A single text file titled "About me" holding your preferences, current projects and standing instructions takes five minutes to write, and it works with any assistant, which beats depending on any company's export. Update it when you notice the assistant has learned something worth keeping.

Then, every few months, do three things. Run your service's data export and store the file somewhere safe, remembering it may contain very personal material. Copy the memory page into your text file. And if you use a self-hosted agent, include its memory folder in your normal backups, leaving out any folder that holds keys or passwords.

If you are in Quebec and want more than the tools offer, you can send a portability request to the company's privacy contact, asking for the information it collected from you in a structured format. The response will tell you a good deal about how the company sees your data, whatever it ends up handing over.