Open ChatGPT right now. Click your name in the bottom-left, then Settings, then Personalization, then Memory. There's a button that says "Manage memories."
Click it.
Now read what's there.
If you've used ChatGPT for the past year or so, you're probably about to have a slightly uncomfortable experience. You're going to see, written out in plain English, a profile of yourself that ChatGPT has been quietly assembling. It will likely include things like: your job, your relationship status, your goals, your insecurities, the projects you're working on, the names of people in your life, the medical conditions you've mentioned, the things you've been worried about lately, your personality, your political leanings, the books you've read, the food you don't like.
Some of it you intentionally told it. Most of it you mentioned in passing, while asking about something else, and ChatGPT decided to remember.
That's the memory feature. It rolled out as a research preview in February 2024, became generally available in September 2024, and was upgraded in April 2025 so it could reference all of your past conversations, not just the ones you explicitly asked it to remember. By June 2025, it was on the free tier. Most users never opted in. Most users still don't know it's on.
ChatGPT isn't alone. Claude added persistent memory in late 2025 and early 2026. Google Gemini has had what it calls "Personal Context" — building from your Gmail, calendar, search history, YouTube — for over a year. By March 2026, all three platforms also had cross-platform memory import: you can now move your ChatGPT profile into Claude, or vice versa, with a few clicks.
Memory is no longer a feature. Memory is the product.
What automatic memory actually builds
When the major AI tools market memory as a feature, they show you the friendly version. "Remembers your preferences." "Recalls past conversations." "Learns your style."
What it actually builds is a profile. Not a list of preferences — a psychological model of you, derived from things you said while distracted, while venting, while half-asleep at 2am, while working through something hard. Most people would not consciously hand over the contents of that profile if asked directly. They handed it over a sentence at a time, over months, while asking about other things.
A pattern I've seen repeatedly: someone mentions a health concern in passing while asking about something unrelated. ChatGPT, doing exactly what it's designed to do, files that health concern into the user's profile. Months later, the user is asking about a totally different topic, and the model's response is subtly shaped by the fact that it "knows" something about their health they'd long since forgotten they mentioned.
That's not a malfunction. That's the feature working as designed.
The settings most people don't touch
Every major AI tool has settings to control memory. Almost no one uses them. Here's the actual map, as of May 2026:
ChatGPT: Settings → Personalization → Memory. You can view the full list of stored memories, delete individual entries, or clear all memory. Separately, in Settings → Data Controls, you can turn off "Improve the model for everyone" — this stops your conversations from being used to train the next ChatGPT, but does not prevent the memory feature from continuing to build a profile of you.
Claude: Settings → Capabilities → Memory. Anthropic stores memory in human-readable markdown files you can actually open and read, which is more transparent than the alternative. You can pause memory per-session with Incognito Chat (look for the ghost icon). You can also see imported memories from other platforms if you ever migrated.
Gemini: myactivity.google.com → Gemini Apps Activity. This is where Google logs the conversations and uses them to build Personal Context. You can pause activity, change the auto-delete window from 18 months to 3 months, or turn it off entirely. Note: even with activity off, conversations may still be sent to human reviewers and retained for up to three years per Google's policies.
If you've never touched these settings, they're all on, by default. The profile has been building since you started using the tool.
Why it's a privacy issue, not just a creepy issue
It's tempting to file this under "creepy but harmless." It's not harmless, and the reason is structural.
A profile assembled by an AI system is more concentrated than the same information scattered across your life. Your therapist knows about your relationship. Your doctor knows about your health. Your accountant knows about your money. None of them know all three. ChatGPT, if you've used it for personal things over the past year, probably knows all three plus your job stress, your political leanings, your kids' names, and the email you've been afraid to send.
That concentration is the issue. Once it exists in a single database, it can be:
- Subpoenaed in litigation — the deletion question covers what's already happening here.
- Accessed by employees of the company. Most major AI companies have human reviewers who can read conversations under various circumstances.
- Leaked in a breach. A 2025 incident exposed approximately 300 million AI chat messages across multiple providers in a single event.
- Used to train future models, in ways the original users would never have anticipated.
- Sold or licensed if the company changes ownership, files for bankruptcy, or pivots its business model.
Each of those is a low-probability event in any given month. Compound them across years and across hundreds of millions of users, and the probability that some version of the worst-case scenario plays out, for some significant fraction of users, approaches certainty.
What "memory done right" could look like
I'm not against AI memory. A thinking partner that remembers your context is genuinely more useful than one that doesn't. The question is whether memory is built in a way that puts the user in control, or in a way that quietly assembles a database the user can't see and didn't ask for.
A few principles I think matter:
Memory should be visible. You should be able to read every single thing the AI remembers about you, in plain English, at any time. Anthropic's choice to store Claude's memory as human-readable markdown files is the right shape of this. OpenAI's "managed memories" view is okay; Gemini's is much harder to access.
Memory should be local where possible. The strongest privacy posture is one where memory lives on your device, not in the company's database. This is technically harder but architecturally cleaner.
Memory should be deletable for real. Not "removed from the visible UI while we keep a copy." Actually deleted, from every system that has it.
Memory should default off, not on. Building a psychological profile of millions of users by default — and then making them dig through settings to turn it off — is the dark pattern of our era.
The practical guide to using AI better covers the settings to change today. It takes about 90 seconds per platform.
What we built
We built Blob on a different model. Conversations are encrypted end-to-end. We can't read them, can't sell them, can't be ordered to produce them. The memory we have of you is the memory you choose to give us — not a profile we assembled in the background while you were asking about something else.
Try Blob free. No credit card.
Then go open ChatGPT and read your profile. It's your data. You should at least know what's in it.