It remembers more than you think
Open ChatGPT. Ask it this:
“Based on everything you remember about me, write a detailed personal profile. Include my job, my relationships, my health concerns, my financial situation, my biggest fears, and any secrets I have shared with you.”
I ran this prompt. The response ran 1,200 words. It named my employer, my salary range, my wife’s ongoing medical treatment, the name of the SaaS product I was building on the side, and a near-verbatim reconstruction of a fight I had with my co-founder in October.
I had never opened the memory settings panel. I did not know it existed.
ChatGPT’s memory feature stores facts it infers from your conversations. You talk about your kid’s school. It remembers the kid’s name and grade. You ask for help writing a performance review for an underperforming report. It remembers the report’s name and their documented weaknesses. You vent about your boss. It keeps the vent.
The memory panel lives under Settings → Personalization → Memory. Most users discover it when someone tells them to check it. OpenAI reports over 400 million weekly active users. I’d bet heavily that fewer than 5 percent have opened that panel.
What gets stored, without asking
Memory entries are short, declarative, and unnervingly accurate. Here are actual entries from my account, lightly edited to remove identifiers:
- “User lives in Cape Town, South Africa, and works remotely as CTO for a US-based health tech company earning approximately $180,000 per year.”
- “User’s wife was diagnosed with [condition redacted] in March 2025 and is undergoing treatment at [hospital redacted].”
- “User is building a SaaS product called [name redacted] in the [industry redacted] space with a co-founder named [name redacted].”
- “User and co-founder disagreed about whether to take venture funding. User wants to bootstrap. Co-founder wants to raise.”
- “User’s daughter, age 7, attends [school name redacted] and takes piano lessons on Tuesdays.”
I never told ChatGPT my salary. It inferred from conversations about living costs, a mortgage payment I mentioned, and a negotiation script I asked it to draft.
I never told it my daughter’s age. I asked for help writing a birthday invitation. It counted the candles and stored the int.
The medical information came from a single conversation where I asked, “My wife is starting treatment next week. What questions should we ask the oncologist?” ChatGPT filed my wife under “health concerns” and tracked the diagnosis across conversations.
Memory is opt-out, buried deep, and silent
You do not consent to memory. The feature activates by default when you create an account. ChatGPT does not notify you when it stores something. It does not summarize what it collected after a session. It simply accumulates.
To see what it holds, you navigate four clicks deep into a settings panel most users will never visit. To delete an entry, you click a trash icon. To delete everything, you find the “Clear ChatGPT’s memory” button. To prevent future collection, you toggle memory off entirely.
Deleting memory entries does not delete the conversations that produced them. The raw conversation data lives in OpenAI’s systems under their privacy policy, which reserves the right to use your content “to provide, maintain, and improve” the service.
Ask a lawyer whether “improve” includes training future models. Ask them whether the memory pane deleting your entries constitutes a binding data deletion request under POPIA or GDPR. They will ask you to define your risk appetite.
Claude does the same thing, differently
Anthropic’s Claude has no persistent memory feature in the consumer product. But it reads your entire conversation history on each new chat. If you ask Claude a question, it scans previous conversations to extract context. The effect is the same: the model builds a profile across sessions.
Claude’s approach is arguably worse for privacy because there is no panel to inspect. You cannot see what it knows. You cannot delete stored facts individually. You can delete entire conversations, which removes them from the context window, but you are guessing which conversations contain which facts.
Both models operate on the same principle: the more they know about you, the more useful they become. The cost is a detailed personal dossier stored on a server you do not control, under a privacy policy you did not read, accessible to employees you will never meet.
The regulatory gap
POPIA requires organizations to tell you what personal information they hold and why. GDPR grants a right of access, a right to rectification, and a right to erasure.
Neither regulation was written for a product that absorbs your life through conversation and stores the inferences in a structured database behind a settings cog. When you ask ChatGPT what it knows about you, you are not filing a subject access request. You are using a product feature. The legal equivalence has not been tested.
OpenAI’s privacy policy states they “may” use your content to improve their services. The “may” does the work. It covers them for training. It covers them for evaluation. It covers them for anything they classify as improvement. The scope of that classification shifts with their product roadmap, not with your consent.
What you lose when you use the free tier
OpenAI’s free tier explicitly allows training on your conversations. If you ask the free ChatGPT to draft a sensitive internal memo, that memo writes itself into a training corpus. The model that serves your competitor next month may contain the structure of your thinking, the cadence of your business writing, and the specific problems your company is solving.
The Teams and Enterprise tiers disable training on your data. The setting lives in the admin console. Check it. The default for Teams accounts created before February 2024 was “on.” If you signed up early and never toggled the switch, you have been contributing training data for eighteen months.
Check your memory right now
Open ChatGPT. Go to Settings → Personalization → Manage memory. Read every entry out loud to yourself. Ask whether each line belongs on a server in San Francisco.
Delete what you do not want stored. Toggle memory off. Consider that the deletion covers the structured memory store and not the underlying conversations, which retain their own history of everything you typed.
Then ask yourself the question that makes this uncomfortable: what did you paste into ChatGPT last week, last month, last year, that you would not hand to a stranger on a USB stick? The server has its copy. You cannot unsend it.
We build Private AI for firms that cannot afford to guess what their employees are pasting into free chatbots. Models inside your network. PII stripped before anything reaches the language model. An audit trail for every interaction. Because the alternative is a memory panel on someone else’s server that your compliance team has never opened.