What You Actually Own When Your AI Remembers: Memory, Export and Portability in 2026

Memory stopped being a feature this year. It became the default.

ChatGPT keeps a memory summary and carries it into new chats. Claude remembers across conversations and keeps each project separate. Gemini learns from your past chats. Microsoft 365 Copilot builds memories from your chat history, your saved notes and your custom instructions.

So the sales question is settled. Every serious assistant remembers. The question I could not answer was a different one. If I spend a year teaching one of these things how my business works, what do I actually hold at the end of it.

I read the vendor documentation for all four, plus the release notes, plus the text of the European portability right. Sources read on 21 August 2026. Every claim below has a link and a date.

Four products, four different shapes of memory. The names are similar and the mechanics are not.

What the four assistants remember, snapshot 21 Aug 2026

Two things stand out. Only one vendor tells you where the memory is physically kept. And that same vendor is the one where memory is on unless somebody turns it off. Microsoft Learn says no action is required to turn Copilot memory on, only to turn it off. The rollout message adds that existing admin and user settings are respected, and that if Memory is toggled off it stays off.

Microsoft is the odd one out, and in a good way. Its documentation says memories are stored in the user's Exchange mailbox in a hidden folder, so they follow the same security and compliance policies as other mailbox data. There is even a folder name and an item class so administrators can search for them.

That is a real answer. It means the memory lands in storage the customer already pays for and already governs.

It is not a complete answer. The same page says retention policies and retention labels configured by organisation admins do not apply to Copilot memory, and that memory and personalisation actions do not generate audit log entries. So the memory sits inside your walls, while your usual rules for keeping and logging things do not reach it.

For the other three, the help pages I read do not say where memory lives. That is not an accusation. It is a gap, and gaps matter when you are working out what you own.

Here is where the picture gets uncomfortable.

OpenAI's export page describes the download only as your chat history and other relevant account data. It does not name memory, saved memories, the memory summary or custom instructions. That is silence, not exclusion, and the page should not be read as either a promise or a denial. What the memory page does say is that to fully delete something ChatGPT may know about you, you have to delete every source where it appears, including past chats, archived chats, files and the memory summary.

Anthropic documents two ways to see what Claude remembers. There is a viewer under Settings, Capabilities. And you can ask Claude in chat to write its memories out verbatim, so you can copy them into a file yourself. Neither one hands you a downloadable file.

Microsoft Learn, on the Copilot personalisation and memory page, says saved and inferred memories are discoverable using eDiscovery and Microsoft Graph Explorer, and that custom instructions can be exported by the user through Settings, Personalisation. The end user support page offers viewing, deleting one memory, deleting all memories, and a Saved memories toggle, and no export at all. So the memory is reachable, but through a compliance search run by IT, not through a button you can press.

Google's personalisation help page says nothing about export.

Not one of the four documents a button that hands you a file. For a person acting alone, the closest thing to a documented route is to ask the model what it knows about you and keep the reply.

The switching tools that launched this spring make the same point louder.

On 2 March 2026 Anthropic made memory free and shipped an import tool. It works by giving you a prompt to run inside ChatGPT, Gemini or Copilot, and an input box in Claude's settings to paste the answer into.

On 26 March 2026 Google shipped its own version. Chats move as a file: you upload a ZIP export, up to 5 GB, up to five files a day, and re-uploading overwrites what you imported before. Memory does not move as a file. Memory moves by running a Google-supplied prompt in your old assistant and pasting the summary into Gemini.

Anthropic is honest about how well this works. Its own page says memory imports are experimental and still in active development, and that at this stage Claude may not always successfully incorporate imported memories.

The industry standard way to move your AI memory in 2026 is to ask the model what it remembers, and paste the answer somewhere else.

That is not a transfer. It is a summary you asked for, not a copy you were handed. You ask a system to describe itself, under no obligation to be complete, and then you treat the description as the asset.

Europe has the strongest rule anywhere, so it makes a fair test.

Article 20 of the GDPR gives you the right to receive personal data concerning you, which you provided to a controller, in a structured, commonly used and machine-readable format, and to send it to another controller without hindrance. It applies where the processing rests on consent or a contract and is carried out by automated means. Paragraph 2 adds a right to have the data sent straight from one controller to another, where technically feasible.

Two limits do most of the work here.

The first is the phrase "provided by". The regulators' guidance, adopted in December 2016 and revised in April 2017, is generous about it. Data you knowingly typed in counts, and so does data observed from your use of the service, such as activity logs and search history. But it draws a line at what the company works out about you. It says inferred data and derived data are created by the controller on the basis of the data provided by the data subject, and that a controller can exclude those inferred data.

A memory summary is a description of you that the vendor wrote. Read against that guidance, it sits much closer to derived data than to something you provided. I am not a lawyer, and I could not find a regulator that has ruled on chat memory specifically, so treat this as an open question rather than a settled one. It is still not a comfortable question.

The second limit is plainer. The same guidance says the right should not create an obligation for controllers to adopt or maintain processing systems that are technically compatible with each other.

So the law can get you a copy. It cannot get you a format that the next vendor has to read.

There is an irony in the switching tools worth sitting with.

Google's import feature is reported as not available in the European Economic Area, Switzerland or the United Kingdom. One report also lists business accounts, enterprise accounts and accounts held by under-18s as excluded.

The region with the strongest portability law in the world is the region where the switching tool is not offered. I found no first-party explanation, and I am not going to invent one. It may be caution, it may be sequencing, it may be something else. What it is not is a good look for portability as a promise.

The lesson is narrow and useful. A feature that helps you leave is still a feature, and features get withdrawn, delayed and limited by region like any other.

Memory is not a vault. It is a product decision, and product decisions move.

Six months of memory changes at one vendor, read on 21 Aug 2026

The June change is the interesting one. A list you curated by hand became a summary the system maintains for you. That is defensible: automatic memory stays current and costs you no effort. It is also a quiet swap of a record you controlled for a record the system controls.

There is one more entry I want to include carefully. In November 2025 a technology outlet reported users finding their saved memory panel blank, with one person describing the whole panel emptying with no history left in it. There was no acknowledgement on the status page, so it was never confirmed as an incident and may have touched very few people. I include it because the advice in that article is the tell: back up your memory by copying it into a text file. That is the state of the art, and it is a manual copy and paste.

Ownership is not one thing. It breaks into questions that have different answers.

What ownership means in practice, snapshot 21 Aug 2026

Look at the first two rows, then the next two. Seeing and deleting are solved everywhere. Holding a copy, and keeping it somewhere you already control, are solved almost nowhere.

That gap is the whole article. You have excellent control over memory while you stay. You have very little claim on it when you go.

The evidence I found most convincing comes from a vendor rather than from me.

Anthropic sells a chat product where memory is a service you can view and edit but not download. It also sells a coding tool where memory is a set of plain markdown files sitting in named folders on your own machine. The documentation for the second one says those memory files are plain markdown you can edit or delete at any time, and that the memory is machine-local.

Same company, two products, two opposite answers. In the assistant, memory is something the vendor holds for you. In the developer tool, memory is a file on your disk with a path you can type.

Only one of those moves when you change engines.

Four questions to ask about any memory feature

None of those questions is about which assistant is best. They are about which parts of what your assistant knows would still exist on a morning when the assistant does not.

I should say plainly that I am not neutral. I teach an eight-week programme where people build their own business brain, so a conclusion that says "keep your own files" is a conclusion that suits me. Read section 09 rather than trusting me. The method I teach is called Arete AIOS (AI Operating System), and what it produces is ordinary text files in a repository the owner controls.

I run my own business that way. The rules for how it writes, the facts about clients and prices, the jobs I repeat, and the decisions I have already made with the reasons attached, all sit in files. The assistant reads them at the start of the work, inside the permissions I set. When I change engines, I change the reader, not the library.

I also want to be fair to the vendors, because files are not magic. A file nobody maintains goes stale, and a stale file is worse than a memory that at least keeps up with you. Vendor memory is genuinely better at one thing: noticing what you never thought to write down.

So the honest position is not files instead of memory. It is files underneath memory. Let the platform notice things. Keep the things that matter where you can open them.

Recall is what the vendors sell. The summarising, the ranking, the noticing, the getting better at it every few weeks. That is real work, it is worth paying for, and it will keep changing on their schedule rather than yours.

The canon is the other half. The written rules, the facts about your business, the decisions with reasons attached. That half has no reason to live inside a chat product, and every reason to live in files you can open, so that any engine can read them and no engine can keep them.

A business brain that exists only inside one vendor is not a business brain. It is an account.

So run the test below on your own business. Eight things your assistant knows. For each one, say where it actually lives. It takes about two minutes, it runs in your browser, and nothing you click is sent anywhere.

If more than half of your answers are "only in the assistant", you already know what you own.