What AI memory means for privacy
An assistant that remembers you can save time. It can also carry personal information into contexts where you did not expect it to appear—and deleting what it remembers may take more than one step.
2026-11-12 · 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.

AI memory sounds like a simple promise: tell an assistant something once, and you will not have to repeat it. The privacy question is less simple. If information from one exchange shapes another, it matters what was kept, where it came from, and whether you can remove it.
There is no single mechanism called “memory.” An assistant might retrieve a stored note, consult an earlier chat or use information from a file or connected service. That is different from information a model may absorb during training. OpenAI, for example, says ChatGPT can draw on saved memories and relevant past chats alongside other available sources, and that the information it uses can change over time. Its features and controls also vary by plan, region, platform and workspace. OpenAI’s memory FAQ describes those distinctions.
What changed: a conversation need not stay in its conversation
Without cross-chat personalization, a detail shared to answer one question need not inform the next. Memory changes that boundary. A preference remembered for later can be helpful: the assistant has more context and may ask fewer repetitive questions. But a detail offered for a particular task can also turn up in a different one. The concern is not just storage; it is reuse.
That distinction matters when judging whether reuse is appropriate. The UK Information Commissioner’s Office (ICO) says organisations using AI should consider why they process personal information and whether a new use fits people’s reasonable expectations. An agent’s ability to identify another use does not, by itself, make that use compatible with the original purpose. This is guidance about purpose and fairness, not a claim that every remembered detail is misused.
The boundary can also be hard for a user to see. OpenAI says ChatGPT’s visible memory summary is high-level and may omit information. It says memory does not retain every detail, either. Both points matter: a summary can help someone inspect personalization without being a complete account of every source that might inform a reply. OpenAI’s memory FAQ is the place to check its current product-specific controls.
Why small details add up
Privacy is not limited to secrets typed explicitly into a chat. Several ordinary details, considered together, may point toward something a person never meant to disclose. The US National Institute of Standards and Technology (NIST) identifies the inference of identities or previously private information as an AI privacy risk. Its generative-AI profile also notes that systems can generate or correctly infer sensitive information.
What does that look like in memory research? One study of people’s perceptions describes concerns about persistent profiles built from information extracted across conversations. It interviewed 18 users. That makes it useful evidence about expectations and concerns, not a measure of how often a particular harm occurs across AI products.
A separate preprint on persistent memory tested simulated, multi-session support conversations involving 24 personal attributes. Under those conditions, consulting memory improved response quality but also increased how many attributes could be inferred from the accumulated record; the gap grew across sessions. The authors describe a possible feedback loop in which the assistant brings earlier information forward and the user affirms, or lets stand, what it assumes. Simulated conversations show a plausible pathway, not a universal outcome for real users.
Together, these findings frame the trade-off more precisely than “memory is good” or “memory is bad.” More context can help an assistant respond well. It can also give the assistant—and anyone evaluating its output—more material from which to draw conclusions about a person. How often that creates a privacy problem in everyday use remains unsettled.
The deletion problem has several parts
A memory setting is not necessarily a delete button for everything a user has shared. In ChatGPT, OpenAI says saved memories are stored separately from chat history. Deleting a chat alone does not necessarily remove a saved memory; deleting that memory does not erase its mention from earlier chats. Its memory FAQ says fully removing remembered information may require deleting the saved memory and the chat where it was shared, as well as addressing other sources containing it.
Timing matters too. OpenAI says it may keep logs of deleted saved memories for up to 30 days for safety and debugging. Turning off reference to chat history schedules information derived from past chats for deletion within 30 days, while the original chats remain unless separately deleted. Those are OpenAI’s published terms for ChatGPT, not a rule for every assistant. They illustrate why “turn it off,” “stop using it” and “erase it” should be treated as different requests. Source: OpenAI’s memory FAQ.
Memory is not the same question as training
Users also need to distinguish personalization from model improvement. OpenAI says its controls for whether personal-account conversations and remembered information may be used to improve models are separate from memory controls. It says Business, Enterprise, Edu and Healthcare workspace content is not used for training by default. Turning memory off should therefore not be read as an answer to every question about data use or retention. OpenAI’s memory FAQ and data-controls page explain its current distinctions.
Temporary mode is another choice, not a magic eraser. OpenAI says temporary ChatGPT conversations do not create or update memories, while limited safety-related use and retention may still apply. Before relying on any provider’s temporary or non-personalized mode, check what it says about saving conversations, safety processing and deletion. Source: OpenAI’s memory FAQ.
What it means for users—and what remains unknown
The most useful privacy check is a set of separate questions: What information can this assistant draw on? What is stored for later? Can I inspect it? Which setting stops future personalization? Which action deletes the underlying material? Checking one box may not answer the others.
A practical starting point is to share only what the task needs, particularly when health, financial, identity or confidential workplace information is involved. That principle aligns with the ICO’s data-minimisation guidance: organisations should process only personal data necessary for their stated purpose. Users can also review available memory controls and check chats, files and connected sources separately when trying to remove information. For workplace or regulated information, organisational rules deserve their own check; product controls may differ across workspaces. Source: OpenAI’s memory FAQ.
The limits of today’s evidence are important. The cited interview study is small, and the multi-session experiment used simulations; neither establishes a privacy-harm rate for everyday use. Product documentation explains published behavior but cannot tell an individual exactly which details an assistant will use in every future response. The defensible conclusion is narrower: memory creates more opportunities for useful continuity and for unexpected reuse or inference. Privacy depends on the sources an assistant can consult, the purposes for which it uses them, and whether its controls match what people reasonably think those controls do.
Our take (opinion)
Our view: the best test of an AI memory feature is not whether it can recall a fact. It is whether a person can understand the route that fact took into a later answer—and stop that route when they want to.
OpenAI deserves credit for documenting distinctions that are easy to blur, including saved memories versus chats and memory controls versus model-improvement controls. The same documentation shows why a single reassuring toggle is not enough. People should not have to guess whether stopping personalization also removes a stored note, its source conversation or a copy elsewhere.
We would judge providers by the clarity of those boundaries, not by how much an assistant remembers. Research suggests a real usefulness–privacy tension, but simulated conversations and a small interview study are not grounds for claiming widespread harm. The case for better inspection and deletion controls does not require that claim.
Questions
Does turning off AI memory delete my old chats?
Not necessarily. OpenAI says turning off ChatGPT’s memory features does not by itself delete original chats. Check the provider’s separate chat-history and deletion controls. https://help.openai.com/en/articles/8590148-memory-faq%C2%A0
Is an AI remembering my preferences the same as training on my conversations?
No. Personalization can use stored information or past chats, while model improvement is a separate process with separate controls. https://help.openai.com/en/articles/8590148-memory-faq%C2%A0
Can AI memory reveal something I never explicitly shared?
It can create an opportunity for inference from details accumulated across exchanges. NIST identifies such inference as a privacy risk, while the cited memory experiment demonstrates a possible pathway under simulated conditions—not its frequency in everyday use. https://nvlpubs.nist.gov/nistpubs/ai/nist.ai.100-1.pdf https://arxiv.org/html/2610.28315v1
What should I check before using a temporary chat?
Check whether it creates memories, how long the provider retains it, and whether safety-related processing still applies. OpenAI says its temporary chats do not create or update memories but may still receive limited safety-related use and retention. https://help.openai.com/en/articles/8590148-memory-faq%C2%A0
Sources
- Memory in ChatGPT | OpenAI Help Center
- Data controls in ChatGPT | OpenAI Help Center
- Theme 5: Fairness and purpose limitations in agentic AI | ICO
- Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile | NIST
- Understanding Users’ Privacy Perceptions Towards LLM’s Retrieval-Based Memory | ACM
- A Self-Reinforcing Privacy Loop in LLM Agents with Persistent Memory | arXiv
- How should we assess security and data minimisation in AI? | ICO
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.