What ChatGPT remembers about you, and how to check
Ask ChatGPT to roast you and it will use what it knows: your job, your city, something you asked about months ago. That is the memory feature working as designed, and it makes a useful audit. Type "Roast me using everything you know about me" and read what it pulls in. Then open Settings, Personalization, Manage memories to see the list itself. Most people are surprised by how much is there, and more surprised that deleting the chat did not delete the memory.
Memory is on, and it saves things you did not ask it to save
OpenAI's documentation is clear that ChatGPT may save details as a memory without you needing to ask. Those saved memories are part of the context for future answers, so a detail you gave once in March can shape an answer in September. This is not hidden. It is a documented feature with a settings page. The gap is that almost nobody opens that page, while almost everybody types real details into the chat box: a client name, a home address, a diagnosis, a salary figure.
The roast prompt closes that gap in about ten seconds, which is why it is worth trying even if you find the trend tiresome.
Deleting the chat does not delete the memory
This is the part that catches people. Saved memories live separately from your conversations. Deleting a chat does not remove the memory it created. Turning memory off does not delete anything either, it just stops new memories forming.
There is a further wrinkle. If you turn memory off and later turn it back on, ChatGPT can create new memories from the chats still sitting in your history, including old ones. So a cleanup that only switches the toggle can quietly undo itself.
To remove something properly you have to delete it in both places: the entry in Manage memories, and the conversation where you first shared it.
How to check and clear it
- Open Settings, then Personalization.
- Open Manage memories and read the list.
- Delete individual entries, or use the memory summary page to delete everything and switch memory off in one action.
- Delete the chats those details came from, including archived ones.
- For one-off questions you do not want retained, use Temporary Chat.
Source: OpenAI's memory FAQ at help.openai.com.
This is not only ChatGPT
Claude, Gemini and most other assistants now have some form of memory or chat history reference. The specifics differ, the shape does not. If you use more than one tool, you have more than one store of personal details to check.
For an individual that is mildly uncomfortable. For a team it is a different question. When staff paste patient details, client records or credentials into an AI tool, retention and memory settings become someone else's policy, not yours. Our write-up on healthcare staff pasting patient data into ChatGPT covers how that plays out in practice.
The part settings cannot fix
Clearing memory deals with what is already stored. It does nothing about the next prompt. That is where PiiBlocker helps: it runs in Chrome and masks names, emails, card numbers, national IDs and other personal data before the prompt reaches ChatGPT, Claude, Gemini, Grok, DeepSeek, Kimi or Perplexity. Detection happens in your browser. Nothing is sent to us. How PiiBlocker works, in technical detail explains the local pipeline, and the Detection settings entry lists the categories it checks.
It is worth being plain about the limits. No masking tool catches everything. An unusual surname can be missed, and nothing can stop someone describing a sensitive situation in plain words. It reduces how much personal data leaves your browser, it does not reduce it to zero.
If you want the settings-only route, the steps above are the whole job, and they cost nothing.