A support lead asks ChatGPT in January to rewrite a difficult complaint. The prompt names the customer, Aoife Walsh, and mentions an open dispute. Months later the same person asks something unrelated, about an internal process. ChatGPT casually refers to "a case like the Aoife Walsh one." The name was never typed again.
A recruiter has Gemini sharpen a job posting and mentions, in passing, that the team just went through a restructure. Weeks later, on an unrelated question, Gemini works that restructure into its answer. Context that belonged to one conversation has bled into another.
This is not a bug and not a breach. This is memory doing exactly what it is designed to do.
What AI memory actually is
Until recently, every AI conversation was a clean slate. You opened a new chat and the tool knew nothing about the last one. That has changed. The major AI tools now remember things between conversations, and they do it in two ways that work independently.
The first part is saved memories: facts you explicitly ask the tool to keep. "Remember that I work in support" or "I always write in British English." The tool stores these as discrete notes and applies them in later chats.
The second part is referencing your past chats: the tool reads across what you typed before and infers preferences and context on its own, without you asking. This is the part most people miss. You do not have to ask the tool to remember anything to be remembered anyway.
The distinction matters. Saved memories can be read back as a list. The context derived from past chats is fuzzier: it is not a single note but a pattern the tool pulls from your history.
What gets remembered, and why that is a work risk
For personal use, memory is mostly convenient. The tool remembers you are vegetarian or that you prefer short answers. At work, the balance shifts.
The data you put in a work prompt is exactly the data you would rather not see resurface months later. A client name in a rewrite request. A project number in a summary. A line about a sick-leave request in a draft letter. A supplier name and an overdue payment in an email you asked it to draft.
When the tool remembers that context, three problems follow. Context bleeds from one conversation into another, so sensitive information from an old case can surface in an unrelated task. Identifiability lasts longer than the employee expects, because the original chat is long forgotten while the derived memory is not. And it is invisible: nobody decides on purpose to keep that client name, it just happens.
Under GDPR, a prompt containing personal data is a processing event. If that same name comes back through memory in a different conversation, it is no longer an isolated event but processing that runs on longer than the organisation assessed. The question "what data is in our AI tools?" gets harder to answer as memory ages.
Per tool: how to view and manage memory
The settings differ by tool. Here are the paths as they currently stand.
ChatGPT
ChatGPT has both kinds of memory. Reference saved memories are the memories you ask it to keep; Reference chat history draws on your earlier conversations. OpenAI rolled out the expanded version in April 2025 and brought it to free users in June 2025 (source: OpenAI Help Center, 2025).
Go to profile icon > Settings > Personalization. There you toggle both controls on or off. Turning off "Reference saved memories" also turns off "Reference chat history." Use Manage memory to see saved memories as a list and delete them individually, or clear everything at once.
Two things surprise people. Turning memory off does not delete what is already remembered; you have to clear that separately (source: OpenAI Help Center, 2025). And deleting a chat does not automatically remove a memory created from it. For a sensitive one-off task, use Temporary Chat: it does not appear in history and does not create memories.
Google Gemini
Gemini personalises responses based on your past chats. Google calls this Personal context and began rolling it out on 13 August 2025. This matters for UK and EU users: the feature started first outside the EEA, UK, and Switzerland, with expansion to those regions in the following weeks (source: Google Blog, August 2025). Check what is on in your own account.
In the Gemini web app, go to Settings, choose Personal context, then the option for your past chats with Gemini, and turn it off or on. To delete something Gemini remembered about you, manage and delete the relevant conversations through Gemini Apps Activity (Menu > Settings & help > Activity). That is also where you set auto-delete, for example after 3 or 36 months. Note that even with activity off, Google keeps conversations for up to 72 hours to provide the service. For a sensitive task without personalisation, use Temporary Chat, which does not appear in your activity and is not used for personalisation.
Microsoft Copilot
Copilot gained a memory and personalisation feature in 2025 that remembers facts about how you work and recurring topics. The control is called Personalization and memory.
In the consumer version on copilot.microsoft.com, go to your profile > Privacy > Personalization; in the desktop app, go to profile icon > Settings > Privacy > Personalization. There you turn it off or manage what has been remembered. The control only appears when you are signed in with a personal Microsoft account and the feature has reached your region through Microsoft's staged rollout.
For Microsoft 365 Copilot, the work version tied to your organisation account, memory and personalisation run through separate user and admin controls; Microsoft scheduled the broader rollout of these for early 2026 (source: Microsoft 365 Message Center / Microsoft Learn, 2025). Here too, turning it off does not automatically delete what has already been collected; that is a separate action.
Claude
Claude has memory and the ability to search past chats. Anthropic made memory more widely available, including the free plan, in 2026 (source: Claude Help Center, 2026). Actively searching past conversations requires a paid plan.
Go to Settings > Capabilities. There you find "Search and reference chats" and "Generate memory from chat history," and View and manage memory shows what Claude has stored. You can delete individual memories, or choose Pause (Claude keeps existing memories but makes no new ones) or Reset (all memories are permanently deleted, which cannot be undone). For a conversation with no memory, use Incognito chat, which stores nothing in memory.
Steps verified in June 2026 with the web apps of ChatGPT (chatgpt.com), Gemini (gemini.google.com), Microsoft Copilot (copilot.microsoft.com), and Claude (claude.ai), and with OpenAI, Google, Microsoft, and Anthropic help and support pages updated in 2025 and 2026. Menu paths and rollout vary by region, account, and version.
What managing memory does not solve
Viewing, limiting, and clearing memory is sensible. It only works after the fact. It does not decide whether the client name, the account number, or the project code should have been in the prompt at all.
A toggle set to off does not erase a memory that is already there. Manage memory cannot see that the name in your next prompt is sensitive. Temporary Chat stops memory from forming, but not the typing and sending of a confidential line. And most employees rarely check their memory settings; they open the tool and get to work.
What remains is the moment of typing. The few seconds when the sensitive data is still on your screen and not yet sent.
How BeeSensible helps before the context is remembered
BeeSensible checks sensitive data in text fields while you type. Through the desktop app, detection runs entirely on the device and no text leaves the machine. For browser-only use, the extension sends the text to BeeSensible's EU detection service, where analysis runs in working memory and the text is discarded after detection. In ChatGPT, Gemini, Copilot, and Claude, names, email addresses, phone numbers, account numbers, payment-card details, and other configured categories get a highlight before you press Enter.
You decide what happens next: remove the value, replace it with a placeholder, or mask it. That way the sensitive context never enters the prompt, and so never enters the tool's memory. BeeSensible does not store the prompt text. Admins see patterns by category and application, without reading employee conversations.
Outlook
