I’ll be honest with you. The first time I heard ChatGPT could now remember our conversations, I thought of that neighbour everyone seems to have. The one who somehow knows your whole business without you ever quite telling them. Bits dropped over the garden fence, added up over months, until one day they ask after your daughter’s exam results and you can’t for the life of you remember mentioning she had any. Clever, in its way. But also, hang on, how do you know that?
That’s roughly where I’ve landed with this. Clever. Useful. And a bit much, all at once.
But here’s the thing. This isn’t a minor update dressed up for a headline. It’s a proper shift in how the tool actually works, more colour telly than yet another camera lens bolted onto your nephew’s phone.
Think about it this way. You know how maddening it is ringing your bank and explaining your whole situation to four different people, starting from scratch every time? AI used to work exactly like that. Every conversation was a blank slate. Tell ChatGPT on Monday you’re vegetarian, and by Tuesday it’s suggesting beef Wellington like the two of you go way back.
Not any more.
Why This Matters More Than You Think
For the first time, we’re dealing with AI that actually pays attention across conversations rather than just within them. It’s the difference between talking to someone who’s genuinely listening and someone who’s just waiting for their turn.
I’ve been using computers since “mouse” meant something that lived in your garden shed, and this one’s earned its place among the handful of updates that actually changed how I use the thing. We’re moving from tools we have to adapt to, towards tools that adapt to us. Whether that’s entirely comfortable is a different question, and we’ll get to it.
What It’s Actually Used For (And What It Definitely Isn’t)
Say you’re writing a novel. In January you mention your protagonist Margaret is terrified of heights. Two months later you’re stuck on a scene and ask for help. With memory switched on, ChatGPT recalls Margaret and her fear of heights without you saying a word. It’s like having a writing partner who actually paid attention the first time.
Same goes for work. Tell it once that you write in British English with Oxford commas and a formal tone, and it remembers. No more pasting the same instructions document into every new chat.
What it isn’t doing, and this matters, is recording your chats to flog you things directly, sharing your details with your neighbours, or quietly building a dossier to sell you a timeshare in Málaga. The stated aim is simply to make your interactions less repetitive. That said, I’m naturally wary of anything that remembers better than I do, and there’s a genuinely sharper edge to this that I’ll come to under the privacy section, because OpenAI’s own recent updates have made it more relevant, not less.
The Before Times: When AI Had the Memory of a Goldfish
Cast your mind back to the early chatbots. You probably don’t need to cast it far, because most of them were useless until fairly recently. I remember a customer service bot in 2018 that couldn’t remember what I’d typed three messages earlier. Talking to it was like talking to my uncle after his third whisky: nodding along, retaining nothing.
The original ChatGPT launched in November 2022 and could write poetry, explain quantum physics, help with homework, all sorts. But every conversation existed in isolation. Close the window and the lot vanished. You could spend three hours on Tuesday explaining your business model, and by Wednesday you were back to square one. Like Groundhog Day, except Bill Murray eventually learned something.
The workaround was tedious. People kept external documents listing their preferences and pasted them in each time. Others just resigned themselves to repeating everything. We adapted, because that’s what we do, but it wasn’t ideal.
The Evolution: From Goldfish to Elephant
Custom Instructions (July 2023)
OpenAI’s first real step towards memory was Custom Instructions, launched in July 2023. Calling it “memory” would be generous. It was closer to leaving yourself a Post-it note: a text box where you could jot down things like “I’m a teacher in Manchester” or “always use metric” or “severe nut allergy, don’t suggest recipes with nuts.” Sensibly, if predictably, it wasn’t available here in the UK at launch. We had to wait while the regulatory to-ing and fro-ing sorted itself out.
It was better than nothing, the way keeping a diary is better than trying to hold everything in your head. But ChatGPT wasn’t learning anything. You were manually programming it, every single time you updated the box.
The First Real Memory (Early 2024)
In February 2024, OpenAI began testing genuine memory with a small group of Free and Plus users. Now ChatGPT could pick up facts from conversations without being told to remember them. Mention you’ve got two daughters in one chat, and weeks later it recalls that detail unprompted.
You could see exactly what it had stored and delete anything you didn’t fancy it keeping, or switch the whole thing off. The control sat with you, which I appreciated, being old enough to remember when technology just did things to you regardless.
Wider Rollout and “Dreaming” (2024–2025)
By September 2024, memory had rolled out properly to Free, Plus, Team and Enterprise users, with a proper notification when it updated and a page where you could review or forget anything.
Then, in April 2025, came the bigger shift. Memory split into two parts: “saved memories” you’d explicitly asked it to remember, and “chat history”, meaning insights it quietly gathered from past conversations generally. The background process that does this sifting has an actual internal name at OpenAI: dreaming. Not a marketing flourish I’ve made up, that’s genuinely what they call it. By June 2025, free users got a lighter version offering short-term continuity, while Plus and Pro users got the fuller long-term picture.
The Current State (2026)
Rather than guess at where things stand now, I went and checked, because the honest answer changes fast enough that guessing felt irresponsible.
OpenAI has just rolled out a new, more efficient memory architecture built on top of dreaming, alongside a “memory summary” page where you can actually see what it thinks it knows about you, correct it, or delete it outright. According to OpenAI’s own internal testing, this update took factual recall from around 68% accurate to 83%, how well it sticks to your stated preferences from 55% to 71%, and how accurate it stays as time passes from 52% to 75%. It’s launched to Plus and Pro users in the US first, with wider countries, including presumably us, following over the coming weeks.
Also worth knowing: ChatGPT can now be connected to Gmail, Google Calendar, Microsoft Teams, Outlook and GitHub, and will reference them in chat where relevant. So “remembering you” increasingly means reading your actual inbox and diary, not just what you’ve typed into the chat box. That’s a genuinely useful convenience. It’s also a bigger ask than it sounds.
How It Actually Works: The Mechanics Behind the Magic
Right, let’s pull back the curtain, and I’ll keep this simpler than your old VCR manual.
Picture yourself at a dinner party. Your conscious mind is having the conversation, responding, asking questions. But underneath, your brain is quietly filing things away: Janet’s daughter just started university, Mark’s allergic to shellfish, Sarah’s off to Scotland in the fashion.
ChatGPT does something similar during “dreaming”, though obviously the wiring is rather different from a human brain. It picks out things that seem significant: personal details, preferences, recurring themes, and stores the essence of them rather than a full transcript of every chat, which would be unwieldy.
When you open a new conversation, it checks what it knows that might be relevant before it replies. Ask about recipes, it might recall you’re vegetarian with a nut allergy. Ask for writing help, it remembers you want British English and a formal tone. It’s not dumping everything it knows into every answer, it’s being selective.
You can explicitly say “remember that I prefer…” or ask what it remembers about you outright. You can also tell it to forget something, or clear the lot. Like having an obliging friend who’ll swear blind they never heard that embarrassing story, if you ask them to.
What’s Coming Next: The Future of AI Memory
Predicting technology is a mug’s game. I once thought MiniDisc players were going to take over the world.
With that caveat firmly in place, the direction of travel seems fairly clear. Memory is likely to stretch across different AI tools and platforms rather than staying locked inside one, so your preferences follow you around instead of you explaining yourself afresh to every app. Systems will probably get sharper at understanding not just what you told them, but when, and how your circumstances have shifted since.
The genuinely tricky questions, ownership, portability, what happens to your data if you switch services or when you’re no longer around to ask, are only going to get louder. I don’t have tidy answers to those. Nobody does yet.
The Serious Bit: Security and Privacy Concerns
I don’t want to sound like the neighbour who thinks the government’s listening through his kettle, but this bit matters.
Memory means ChatGPT is storing information about you. That’s literally the job. Be mindful of what you type into it. Don’t share passwords, financial details, or anything that could help someone steal your identity, and don’t share other people’s private information either, that’s just good manners as much as good sense.
Your memories sit on OpenAI’s servers, not your device, so they’re subject to whatever security OpenAI has in place, and potentially to legal requests for data. If you’re discussing anything sensitive, commercially confidential, or legally protected, think carefully before it goes in.
There’s also the plain risk of a data breach. Any system can be hacked. If someone got into ChatGPT’s memory store, they’d have access to a genuinely huge amount of personal information across millions of users.
You can manage a fair amount of this yourself. Review what’s stored and delete anything you’re not comfortable with. Use temporary chat mode for anything sensitive, since it doesn’t get saved at all. There’s also a Lockdown Mode you can switch on in Settings, which restricts things like live web browsing and file downloads to cut the risk of the AI being tricked into leaking your data through a poisoned webpage or document. Worth knowing it exists, even if you never need it.
If you use ChatGPT for work, particularly in healthcare, law or financial services, check what your industry and employer actually allow before you type anything work-related into it.
And here’s the bit that gave me pause. OpenAI has just started rolling out adverts to Free and Go plan users in the UK specifically. Paid plans stay ad-free. Read that alongside everything above and draw your own conclusions about how a system that knows you better ends up deciding what to show you, and who’s paying for the privilege. I’m not saying anything sinister is afoot. I am saying the more accurately it knows you, the more useful you become to someone else’s bottom line too.
Wrapping This Up
ChatGPT has gone from a clever but forgetful conversational tool to something that genuinely carries context between chats and adapts as your circumstances change. Is it perfect? No. Is it worth a raised eyebrow on privacy grounds? Also yes. Is it genuinely useful? Also also yes, all three at once, which is rather how most decent technology ends up.
Use it thoughtfully. Take the convenience, but keep an eye on what you’re feeding it, and don’t hand over anything you wouldn’t be comfortable seeing again in six months’ time, because there’s a decent chance you will.
Check occasionally what it’s remembering about you. That one’s not optional advice, that’s just good sense.
Walter
Leave a Reply