Who regulates AI memory? Start with the data, not the label
There is no single rule for everything an AI service calls “memory.” Privacy laws, AI rules and product controls address different parts of the problem—and deleting a chat may leave other records untouched.
2026-11-08 · The Konshus Team · Sources checked before publishing
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.

What changed: The rules are becoming clearer, but not simpler
Ask whether AI memory is regulated, and the answer depends on what is being remembered. A saved preference used in future chats, a searchable conversation history and information absorbed during model training can all be called “memory.” They do not necessarily sit in the same place or respond to the same controls. OpenAI’s memory documentation distinguishes saved memories from chats; the European Data Protection Board’s guidance addresses a different question: personal data in AI models.
There is no single law for every feature bearing the memory label. The more useful starting point is what personal information a service collects, retains and uses, and what a person can do about it. In Europe, the GDPR sets rules for processing personal data, including limits on purpose, amount and retention. Organizations should set time limits for erasing or reviewing stored data, according to the European Commission. A product team cannot settle those questions simply by choosing a different name for a feature.
This is a state-of-play story rather than the arrival of an AI-memory statute. The EU AI Act’s main application began on August 2, 2026, according to the Commission’s timeline. California’s updated privacy regulations took effect on January 1, 2026, while requirements for certain automated decisions are due to begin applying on January 1, 2027, the California Privacy Protection Agency says. Neither milestone means every chatbot memory is governed by the same new rule.
Why the distinction matters
Think of a request to “forget” a personal detail. The detail might appear in an old conversation, a separate saved memory or information used to develop a model. A button that removes one copy does not, by itself, tell you what happened to the others. That is why a useful privacy question is not merely Can I turn memory off? It is Which information will stop being used, and which records will be deleted?
The GDPR gives people rights that include access, correction and, under specified conditions, erasure. Erasure is not absolute: the European Commission identifies exceptions, including legal obligations and legal claims. So it would be misleading to promise that every valid deletion request makes every trace disappear. It would be equally misleading to treat a product’s delete button as the full measure of someone’s legal rights.
Model development makes the question harder. In its Opinion 28/2024, the European Data Protection Board says an AI model trained on personal data is not automatically anonymous. That assessment depends, among other things, on whether personal information can be extracted or obtained through queries using reasonably likely means. The board also says a provider relying on “legitimate interests” to process personal data needs a case-specific assessment of its interest, the necessity of processing and the balance against people’s rights. These are tests to apply, not blanket permission to keep information indefinitely.
The same opinion warns that unlawful processing during model development can affect whether a model may lawfully be deployed later, unless it has been duly anonymised. It discusses ways for people to raise claims that their data has been memorised or reproduced. But it does not establish one universal technical procedure for removing every piece of personal information that has influenced a model. The unsettled issue is how an organization demonstrates that a particular remedy works for the particular data and system involved.
Where the EU AI Act and California fit
The EU AI Act complements rather than replaces data-protection law. Its text says privacy and personal-data protection must be guaranteed throughout an AI system’s life cycle, and refers to data minimisation and protection by design and by default under EU data-protection law. Its staged implementation matters for AI providers, but an AI Act date is not a general authorization to retain someone’s personal history.
California approaches the issue through consumer privacy rights and limits on data use. Its CCPA statute requires covered businesses’ collection, use, retention and sharing of personal information to be reasonably necessary and proportionate to the stated or compatible purpose. Consumers can request deletion, subject to exceptions, according to the state privacy agency. That makes the purpose of a persistent profile important: information collected to answer one question is not automatically needed for every future interaction.
California’s automated-decisionmaking rules address a narrower situation. The privacy agency says its requirements for automated decision-making technology used in significant decisions begin applying on January 1, 2027. An ordinary conversational memory should not be described as automatically subject to those requirements. The analysis can change when personal information is used in a consequential decision rather than simply to personalize a reply.
In the United States more broadly, NIST’s generative-AI risk framework identifies risks involving personal data used in training, including insufficient transparency and the possibility of leaking or inferring sensitive information. It is a risk-management framework, not a general legal right to delete a chatbot’s memory. It helps explain the risk; applicable law determines the obligation.
What the major services actually offer
Product controls are worth examining, but these are providers’ descriptions, not independent verification of what happens inside their systems. OpenAI says that removing information saved as a ChatGPT memory requires deleting both the saved memory and the chat where the information was first shared. Its memory FAQ also distinguishes memory from chat history. That is a concrete warning against assuming that deleting a conversation and deleting a memory are interchangeable.
Anthropic describes a similar separation for Claude: deleting a conversation does not remove memory entries generated from it, which can be deleted separately while memory is on, according to its memory guidance. Its retention guidance says deleted conversations disappear from chat history immediately and are deleted from back-end storage within 30 days. A disappearance from the interface and deletion from storage are therefore distinct events even in the provider’s own account.
Google says Gemini users can review, delete or turn off Apps activity, with activity older than 18 months auto-deleted by default and other periods available. Its activity guidance says conversations can remain with the account for up to 72 hours even when activity is off. Separately, its privacy hub describes saved instructions and a memory summary that users can view, edit or delete. These are useful controls, but their different names and time frames make it worth checking exactly which store a setting affects.
What it means for a request to forget
For users, the practical lesson is to identify the information first, then check each relevant control: saved memory, originating chat and activity history may require different actions. If a concern is about personal data in a trained model, a conversational memory toggle may not address it. If a concern is about future personalization, deleting a chat alone may not address it either.
For organizations, the harder work is mapping where personal information goes and why it remains there. A clear interface helps, but it cannot substitute for decisions about purpose, retention and responses to rights requests. The open question is not whether AI can remember. It is whether people and regulators can tell what has been kept, where it has gone and what “forget” actually changes.
Our take (opinion)
The most promising development is not a new button. It is the insistence, visible in privacy law and regulatory guidance, that organizations account for personal information across its uses. That is a better test than whether a service markets itself as having “memory.”
The major providers deserve credit for publishing controls that distinguish at least some of these stores. OpenAI and Anthropic, in particular, explicitly warn that deleting a chat and deleting a memory can be different actions; Google offers activity and memory controls of its own. Those explanations help users make better choices. They also reveal the remaining weakness: understanding a provider’s stated controls can take more work than saving the information did.
Our view is that a trustworthy request to forget should tell a person, in plain language, what was removed, what was not, and why. Lawful exceptions and technical limits may exist. Neither should be hidden behind a reassuring but ambiguous word like “deleted.”
Questions
Is there a law specifically for all AI memory?
No single law covers everything providers call AI memory. Rules depend on the personal information involved, its use and the applicable jurisdiction. The GDPR, California privacy law and the EU AI Act address different aspects.
Does deleting a chat delete what an AI remembers about me?
Not necessarily. OpenAI says removing a saved memory requires deleting both that memory and the chat where it was first shared. Anthropic says deleting a Claude conversation does not remove memory entries generated from it.
Can I demand that a model unlearn my personal information?
The GDPR provides an erasure right in qualifying circumstances, subject to exceptions. The European Data Protection Board addresses personal data in AI models, but its guidance does not establish a universal technical method for removing every influence of that data.
Do California’s automated-decisionmaking rules apply to every chatbot memory?
No. The rules address covered automated-decisionmaking uses, including specified significant decisions. They should not be treated as a stand-alone rule for every conversational memory feature.
Sources
- For how long can data be kept and is it necessary to update it? — European Commission
- Information for individuals — European Commission
- Opinion 28/2024 on certain data protection aspects related to AI models — European Data Protection Board
- Full text — EU AI Act
- Navigating the AI Act — European Commission
- California Consumer Privacy Act of 2018 — California Privacy Protection Agency
- Frequently Asked Questions — California Privacy Protection Agency
- California Privacy Protection Agency regulations update
- Memory FAQ — OpenAI
- Claude chat search and memory guidance — Anthropic
- How long do you store my data? — Anthropic
- Manage and delete your activity in Gemini Apps — Google
- Gemini Apps Privacy Hub — Google
- Artificial Intelligence Risk Management Framework: Generative AI Profile — NIST
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.
About Konshus: a private memory Vault that works across ChatGPT, Claude, Gemini and more. Learn more.