Konshus
AI Memory News

Who owns the memories an AI keeps about you?

You may retain rights to what you tell an AI without owning every profile or summary it builds from that information. The more useful question is what you can see, change, delete and take with you.

2026-11-14 · The Konshus Team · Sources checked before publishing

Share

Want more news like this?

  • One short email a week with the biggest changes in AI memory — ChatGPT, Claude, Gemini, new companies and research.
  • Every story is checked against its sources before it goes out.
  • No spam: just the weekly news, no sales emails, and we never sell or share your address.

Free. One email a week, nothing else — no Konshus marketing emails, and we never sell or share your address. Unsubscribe anytime in one click.

The short answer: it depends what you mean by “own”

Suppose you tell an assistant that you prefer short answers. Later, it uses that preference without being reminded. Is the preference yours because you supplied it? The provider’s because its system stored it? Or is “ownership” the wrong question?

There is no single answer that covers every kind of AI memory. A useful starting point is to separate four things: rights in information you submit, control over a product’s saved memory, permission to use information for other purposes, and privacy rights over personal data. You can have some of those rights without having all of them. And a button labeled “delete memory” may address only one place where the information lives.

This is a state-of-play explainer, not a report of a newly settled legal rule. The distinctions matter because memory can turn a passing remark into information an assistant uses again. If you want to know what control you have, the practical details of the feature matter more than a broad claim that your data is “yours.”

What changed when assistants began remembering?

The central shift is from information supplied for one exchange to information available for later responses. That does not mean there is one complete, editable record called your memory. OpenAI says ChatGPT can personalize responses using saved memories, past chats, custom instructions, files and connected apps. Its memory summary may not display everything the product can use. OpenAI’s memory guide therefore makes an important distinction visible: what you can inspect as a saved memory is not necessarily the full set of relevant sources.

A preference can also originate in more than one place. You might state it directly, or an assistant might derive a useful summary from a conversation. If the same detail remains in a chat or file after a saved memory is removed, the question is not just who owns the summary. It is where the underlying information still exists and whether the product can still draw on it.

That leads to three separate tests. Can you see and edit what the assistant remembers? Can you remove the sources it learned from? Can you limit other uses of those sources? A service may offer one control without offering the others in the same place.

What the contracts answer—and what they do not

OpenAI’s terms say that, as between the user and OpenAI and to the extent permitted by law, the user retains ownership rights in their input and owns the output. Anthropic’s consumer terms say users retain whatever rights they have in inputs and receive Anthropic’s rights, if any, in outputs; Anthropic and its providers retain rights in the Claude service itself. (OpenAI’s terms)

Those are meaningful commitments. They also need to be read precisely. Input is what you provided; an assistant’s stored summary or inference may be something the service created from that input. The cited terms do not establish one universal rule for ownership of every such memory across products. It would be misleading to treat an input-ownership clause as proof that you own each derived profile. It would be just as misleading to assume the provider owns your original words merely because it runs the memory system.

“Own” can also suggest powers that a contract clause does not promise. Keeping rights in a sentence you typed does not, by itself, tell you whether a saved summary is visible, whether it can be exported, whether its source chat remains on file or whether a deletion request has exceptions. For those answers, you have to look at the product controls and the applicable privacy rules.

What control looks like in ChatGPT, Claude and Gemini

ChatGPT: OpenAI says users can review or correct available memory information and delete saved memories. But saved memories are stored separately from chat history. Deleting a chat does not automatically delete a saved memory made from it, while deleting the memory does not erase references in past conversations. OpenAI advises removing the information from relevant chats, files or connected apps too. (OpenAI’s memory guide)

The switches also have different effects. Turning off memory does not delete past chats. OpenAI says turning off reference to chat history schedules information remembered from past chats for deletion within 30 days, while the chats remain unless you delete them separately. It may keep logs of deleted saved memories for up to 30 days for safety and debugging. These distinctions make OpenAI’s controls more specific than a single on/off promise, though they also give users more than one place to check. (OpenAI’s memory guide)

Use for model improvement is another question. OpenAI says chats and remembered information in personal accounts may be used to improve models when Improve the model for everyone is on; users can turn that setting off. It says content from Business, Enterprise, Edu and ChatGPT for Healthcare workspaces is not used to train models by default. A managed workspace can also be subject to its organization’s settings and policies. (OpenAI’s memory guide; OpenAI’s data controls guide)

Claude: Anthropic says users can view memory topics, edit or delete them, and reset memory. Deleting a conversation does not necessarily delete memory entries generated from it; those entries must be managed separately. Anthropic says memory data is included in data exports. Those are useful inspection and portability controls, but an export is a copy of data, not a promise that another assistant will interpret it the same way. (Anthropic’s Claude memory guide)

There is also no single Claude storage arrangement to describe. For Anthropic’s API memory tool, the developer’s application performs the file operations and controls where and how memory is stored. In that setup, a person’s route to inspecting or deleting memory may depend on the application built with the tool, not just on Anthropic’s consumer controls. (Anthropic’s memory tool documentation)

Gemini: Google says eligible users, depending on their settings, can review, edit or delete available memory summaries and turn memory on or off. Saved instructions and other memory features need not follow identical rules. Google describes an 18-month maximum for certain remembered items or instructions in one experience; that is not a retention rule for every Gemini feature or account. Google deserves credit for documenting feature-specific controls. The qualification is equally important: check the rules for the feature you actually use. (Gemini Apps Privacy Hub; Google’s memory guidance)

Privacy rights are a different kind of power

Privacy law can give you ways to act even when ownership of a particular derived memory is unclear. The European Commission describes GDPR rights that include access to personal data and, in certain circumstances, erasure and portability. It also makes clear that erasure is not absolute: exceptions can apply. (European Commission guidance)

California’s privacy agency describes rights to know what personal information businesses collect and use and to request deletion, subject to exceptions. California’s rules also address deletion requests involving certain service providers and third parties. These are rights governed by conditions, not a universal command that every trace disappears immediately. (California privacy FAQs; California privacy rights)

Whether a particular AI-generated inference counts as personal data, which organization must answer a request, and whether an exception applies can depend on the facts and jurisdiction. That uncertainty should not obscure the practical point: you do not have to resolve a philosophical ownership dispute before asking what data a service holds or what deletion rights may apply.

What to do with an unsettled answer

Start with the memory the product lets you inspect, but do not stop there. Check the chats, files and connected sources from which a detail could be drawn. Look separately at personalization, model-improvement and workspace settings. If an account belongs to an organization, ask who controls its settings: OpenAI says workspace owners and administrators can manage supported memory settings and permissions. (OpenAI’s memory guide)

Finally, be precise about the outcome you want. Correcting an inaccurate preference, stopping future personalization, deleting a source conversation and requesting a data export are different tasks. The current products provide real controls, and their documentation is worth using. But none of the cited terms or help pages turns “I own my memory” into a guarantee that one action removes every source, inference or retained record. The defensible answer is narrower—and more useful: identify the information, identify where it lives, then use the controls and legal rights that apply to each place.

Our take (opinion)

“Who owns it?” is an understandable question, but it risks producing a reassuring answer that is too vague to help. We think providers should make the boundaries between saved memory, source material and other uses of personal information much easier to see. Users should not need to investigate several settings to learn whether deleting a remembered fact also removes the conversation that supplied it.

The companies here have built worthwhile controls. OpenAI explains why chat and saved memory require separate action; Anthropic offers memory inspection and includes memory in exports; Google documents controls and qualifications for particular features. Those details are more valuable than a sweeping ownership slogan.

Our preferred test is practical: can a person discover what is being used, correct it, stop future use, remove the relevant sources and understand any limits? A good answer to each question would give users more meaningful control than an ambiguous claim of ownership.

Questions

If I own what I type into an AI, do I own its memory about me?

Not necessarily. OpenAI’s and Anthropic’s cited terms address rights in inputs and outputs, but do not establish one universal ownership rule for every summary or inference a memory feature creates.

Does deleting a chat delete what the AI remembered from it?

Do not assume so. OpenAI and Anthropic both describe memory entries that may need to be deleted separately from their source conversations.

Does turning off memory delete my old conversations?

Not in ChatGPT: OpenAI says turning off memory does not delete past chats. Check the specific controls of the service you use.

Can privacy law help if ownership is unclear?

Potentially. GDPR and California privacy rules provide qualifying rights concerning personal data, including access or deletion requests. Their scope and exceptions depend on the jurisdiction and circumstances.

Sources

  1. OpenAI, Memory in ChatGPT
  2. OpenAI, Data Controls FAQ
  3. OpenAI, Terms of Use
  4. Anthropic, Consumer Terms of Service
  5. Anthropic, Use Claude’s chat search and memory to build on previous context
  6. Anthropic, Memory tool
  7. Google, Gemini Apps Privacy Hub
  8. Google, Manage what Gemini remembers
  9. European Commission, Information for individuals
  10. California Privacy Protection Agency, Frequently Asked Questions
  11. California Privacy Protection Agency, Rights under the California Consumer Privacy Act

The AI memory news, once a week.

If you're interested in AI memory, want your own AI to remember you better, or just want to stay up to date, this is the newsletter for you. Every week we comb through the news for the big shifts in AI memory — new companies, new features from ChatGPT, Claude and Gemini, new research and new ideas — and sum up what matters in one short email.

For people who use AI every day, builders and founders, and anyone curious where AI memory is going.

Free. One email a week, nothing else — no Konshus marketing emails, and we never sell or share your address. Unsubscribe anytime in one click.

About Konshus: a private memory Vault that works across ChatGPT, Claude, Gemini and more. Learn more.