Privacy & Control
What AI actually stores about you, and how to control or delete it
What ChatGPT, Claude and Gemini store by default, how to see and delete it, why deleting a chat does not delete a memory, and what UK GDPR does and does not give you.
By AI Twin · 23 July 2026 · 5-minute read
Last verified . AI memory features change quickly; dated claims were checked on this date.
Assistants now remember by default. This post sets out what the three major assistants store, how to see and delete it, the one myth that catches almost everyone, and what your rights under UK GDPR actually deliver in practice.
The short answer
ChatGPT, Claude and Gemini each keep two kinds of record: facts you can see and delete in settings, and an inferred picture built from your history that you can only partly inspect. Deleting a conversation does not delete the memories made from it. UK GDPR gives you real rights over both, with practical limits this post explains plainly.
The three things “memory” can mean, and why it matters for privacy
Working memory is the current conversation and vanishes with it. Saved memories are stored facts you can list and delete. Chat-history reference is an inferred profile assembled from everything you have said. The privacy stakes rise in that order: the first is ephemeral, the second is visible, the third is the one you cannot fully read.
When you think about what an assistant knows, ask which of the three you mean. The controls are different for each.
What each assistant stores by default
As of 23 July 2026: ChatGPT stores saved memories and, where the feature is available, references chat history; the chat-history feature was not rolled out to the UK and EEA at its April 2025 launch. Claude stores memory from chat history for all users, on by default since 2 March 2026, with viewable and editable memory settings and an incognito mode. Gemini stores personal context tied to your Gemini Apps Activity, with auto-delete windows you can set.
Defaults change with plans and regions. The verification date at the top of this page is there for a reason.
How to see, edit and delete what each assistant knows
ChatGPT
Settings, then “Personalization”, then Memory shows saved memories, individually deletable or clearable in full. Settings, then Data Controls covers chat history and export. Temporary Chat runs a conversation that neither uses nor writes memory. Turning off both saved-memory reference and chat-history reference stops new remembering.
Claude
Memory settings show what Claude holds, and entries can be edited or removed. Memory can be disabled. An incognito conversation keeps a discussion out of memory entirely.
Gemini
Gemini Apps Activity lists stored activity, with deletion by item or period and auto-delete windows of three, 18 or 36 months. Personal-context features can be switched off.
The myth: deleting a chat deletes the memory
It does not. A saved memory created during a conversation is a separate record. Delete the conversation and the memory survives; delete the memory and mentions of it in old conversations survive. If you want something gone, delete it in both places: the memory list and the conversation history.
AI memory and UK GDPR: what your rights actually deliver
UK GDPR gives you the right to access what a provider holds about you, to correct it, and to ask for its erasure under Article 17. These rights are real and providers operate request routes for them.
The practical limits are worth stating plainly. Erasure applies to stored records: your account data, conversations, and saved memories. It does not un-train a model. If information influenced model weights during training, deleting your records does not reach into the model itself; that distinction is honest, and any provider claiming otherwise is overclaiming. Second, response times and scope vary by provider and by whether you are a consumer or business user. Third, deletion of a record and expiry of backups are separate steps on separate schedules; a well-run provider will tell you both.
None of this makes the rights hollow. It means the strongest position is held by the person who put less irreversible information in to begin with.
What not to put into an assistant’s memory
Treat an assistant’s memory as a convenience layer, not a vault. Financial account details, passport and identity numbers, other people’s private information, and anything you would not want summarised back to you in front of a stranger belong elsewhere. Client and patient information carries professional obligations that consumer assistant tiers are not designed to meet.
Holding memory yourself, with a record of every change
The alternative to trusting each assistant’s memory is holding one memory of your own and letting assistants read it under your rules. AI Twin is built that way: you put things in deliberately, every entry is typed and reviewable, every change lands in an append-only audit log, and nothing an assistant proposes is saved until you confirm it. The design goal is that the answer to “what does it know about me” is never a mystery, because you wrote the list.
Questions people ask
Is ChatGPT memory GDPR compliant?
UK GDPR rights apply: you can access, correct and request erasure of your stored data through the provider’s privacy routes. Erasure covers stored records, not model weights. Whether any given use is appropriate depends on what you put in; consumer tiers are not designed for confidential professional data.
Does deleting a chat delete the memory?
No. Saved memories are separate records and survive the conversation they came from. Delete both the memory entry and the conversation if you want something fully gone from your visible history.
Can I stop ChatGPT remembering me?
Yes. Turn off saved-memory reference and chat-history reference in Settings, and use Temporary Chat for conversations that should leave no trace. Existing memories stay until you delete them.
How do I delete what AI knows about me?
Delete saved memories in the assistant’s settings, delete the conversations, then use the provider’s privacy request route for account-level erasure under UK GDPR. Expect stored records to be deletable and model weights to be out of scope.
What should I never tell an AI assistant?
Identity numbers, financial credentials, other people’s private details, and anything covered by professional confidentiality. The convenience of memory is real; so is the fact that you are writing a record.
Newsletter
Notes from the build.
AI Twin is open to early users now. We will tell you when wider sign-in opens.
More from the Journal
Foundations
What we mean when we say a personal operating system
AI Twin is a personal operating system for your memory. Here is what that phrase means, what it does not mean, and why we use it deliberately.
8 July 2026 · 4-minute read
Field notes
Your memory, in your language. And a note on getting it right
AI Twin works in more than 100 languages. Here is what that means today, what we are still improving, and an open invitation to help us get the translations right.
22 July 2026 · 4-minute read
Practical
Magic Inbox: how forwarding an email becomes memory
Forward any email to AI Twin and it becomes typed, searchable memory. Here is how Magic Inbox works, and where its limits are.
15 July 2026 · 4-minute read

