Why I Turned Off ChatGPT's Memory Feature

A few weeks ago I opened a brand-new ChatGPT chat to ask something completely unrelated to anything I'd talked about before, and it opened by referencing a deadline I'd mentioned in passing to a different chat, on a different day, about a different project. I hadn't asked it to remember that. I'd forgotten I'd said it at all. That's the moment I went looking for the off switch, and it's taken me a few weeks of actually living without Memory to be sure turning it off was the right call rather than a knee-jerk one.

I want to be upfront about something before I get into it: this isn't a "ChatGPT is spying on you" piece. It's more mundane and, I think, more useful than that — a feature I opted into without really deciding to, that quietly changed how the tool behaved, and that I only noticed once it did something specific enough to be unmissable.

How Memory ended up on in the first place

I don't remember explicitly turning it on. It rolled out as a default at some point, folded into a settings menu I hadn't opened in months, and it does genuinely useful things when it's working the way it's meant to — it remembered that I prefer short first drafts over long ones, that I write for this site in the first person, and a couple of recurring project names, which meant I stopped re-explaining context I'd already given it. For a good while that felt like a straightforward win, the same quiet convenience as a browser that remembers your password. I didn't think about what it actually meant that a chat window somewhere was accumulating a running file on me.

The moment it stopped feeling like convenience

The deadline comment wasn't sensitive. That's actually what unsettled me about it — if it had surfaced something genuinely private I'd at least have understood why I felt uneasy. Instead it surfaced something completely ordinary, in a context where I had no reason to expect it, and the effect was the same one you get when a stranger already knows your name before you've introduced yourself. Not dangerous. Just wrong-footing, in a way that made me realise I had no real sense of what else was sitting in that memory store, or how far back it went.

"It wasn't the sensitivity of what it remembered that bothered me. It was that I had no idea what it had decided was worth keeping, or for how long."

What actually lives in there

ChatGPT's Memory settings do let you review and delete individual entries, and I'll give the feature credit for that — it's not hidden. What I found when I actually opened that screen was a longer list than I expected: work habits, a couple of half-formed personal details I'd mentioned once and never again, preferences I'd genuinely rather it kept, all sitting together with no real hierarchy between "remember this forever" and "I mentioned this once, in passing, and it's not relevant to anything else I'll ever ask you." The tool had no way of telling those apart, because I'd never told it to. It just kept everything by default and let me clean up after the fact, which is a very different thing from asking first.

It's the same shape of problem I ran into writing about the VPN privacy myth I believed until I actually tested it: the risk was never one dramatic leak, it was a slow accumulation of ordinary detail that adds up to something a lot more specific than any single piece of it looks like on its own. A VPN doesn't stop that kind of accumulation if the thing doing the collecting is the app you're typing into directly, not the network underneath it.

What I use instead

I turned off "Reference saved memories" in Settings, which stops the model pulling anything forward between chats — it doesn't touch your existing chat history, which is a separate setting entirely, so I also went through and deleted the memory list itself rather than leaving it sitting there unused. The handful of things that genuinely helped, I kept, but on my terms: a short custom instructions block, written by me, that I can read in full in about fifteen seconds and edit whenever it's wrong. It's a smaller list than what Memory had quietly built, and every line in it is one I put there on purpose. I ended up doing something similar to what I found writing about my month with an AI journaling app — the version of "the AI remembers things about you" that actually sat right with me was the one where I decided what got kept, not the one where the software made that call for me by default.

If you're deciding whether to keep yours on

Open Settings > Personalization and actually read your memory list before deciding either way — most people I've mentioned this to had never looked. If most of what's there is stuff you'd have happily typed into a custom instructions box yourself, you've lost nothing by moving it there on purpose and switching the automatic version off. If some of it surprises you, that's worth sitting with for a minute rather than clicking past. I still use ChatGPT for most of my day-to-day work, alongside the wider AI tools I rely on — this changed how much control I hand it by default, not whether I use it. It's the same instinct that made me switch off Meta AI everywhere it showed up uninvited: I don't mind an AI assistant knowing things about me, as long as I'm the one who decided to tell it. If you want the next one of these before it goes up, join the Brite newsletter.

Questions people actually ask

What does ChatGPT's Memory feature actually do?

It quietly saves details from your conversations — your job, your preferences, ongoing projects, things you've mentioned in passing — and brings them back into later chats without you asking. You can review and delete individual memories in Settings, but most people never open that screen, so it just keeps accumulating.

Is ChatGPT Memory a privacy risk?

It's less a single risk than a slow accumulation of one. Nothing is broadcast publicly, but you're building a standing profile of yourself inside a company's servers that persists across every future chat, including ones you didn't expect it to be relevant to. That's a different shape of risk to a one-off data leak, and it's worth weighing on its own terms rather than dismissing because nothing has "gone wrong" yet.

Can you turn off ChatGPT Memory without losing your chat history?

Yes — they're separate settings. Turning off "Reference saved memories" in Settings > Personalization stops it from carrying details between conversations, but your past chats stay exactly where they were unless you also delete them separately.

What's the alternative to letting ChatGPT remember things for you?

Writing it down somewhere you control and can actually see the whole of at once. I moved the handful of standing facts that genuinely helped — how I like a first draft, what my current projects are — into a short custom instructions block I edit myself, rather than letting the model harvest them silently from conversation.

Do you still use ChatGPT after turning Memory off?

Daily. This isn't a case against the tool — it's a case against handing a feature control it hadn't earned. Every new chat starts clean now, and if it needs to know something about the current task, I tell it, on purpose, every time.