On 4 June, OpenAI started rolling out a memory upgrade it calls “dreaming.” ChatGPT no longer just stores what you tell it — it revises what it knows in the background as time passes, quietly turning “I'm going to Singapore in July” into “I went to Singapore in July 2026” once the date has gone. There's a new memory summary page where you can see, correct and delete what it thinks it knows about you, and the recall numbers are real: OpenAI says factual recall has climbed from 41.5% in 2024 to 82.8% now. It's a better product. It also quietly changes the data-hygiene maths for any business whose team pastes customer or financial details into ChatGPT — because the thing now remembers across conversations and rewrites itself without being asked. Here's the one rule I handed every client this week.
Two years ago, ChatGPT had the memory of a goldfish, and that was oddly reassuring. Every conversation started fresh. Whatever you pasted in vanished when you closed the tab, as far as the next chat was concerned. You could be a bit careless because the slate wiped itself.
That era ended quietly on 4 June. OpenAI began rolling out a system it calls “dreaming” — and the name is doing real work. ChatGPT now runs a background process that learns across your conversations and continuously synthesises what it knows about you, so it always has the freshest version of your context to hand. It doesn't just remember facts; it updates them. The example OpenAI gives is small and telling: tell it you're going to Singapore in July, and after July it revises that to “you went to Singapore.” The memory ages itself.
To OpenAI's credit, they've shipped a transparency surface alongside it — a memory summary page where you can see what it thinks it knows, correct anything wrong, delete what shouldn't be there, and tell it which topics to raise and when. And the capability jump is real: factual recall on their own internal test went from 41.5% in 2024 to 82.8% today. As a product, this is a clear upgrade. As a thing your team uses for work, it deserves about ten minutes of thought you probably haven't given it.
Why a better memory is good
Persistent, self-updating memory makes the assistant dramatically more useful for the repetitive, context-heavy work small businesses do. It stops you re-explaining your business, your tone of voice, your customers and your constraints at the start of every session. For the way I use it — mostly drafting and planning — it removes a tax I'd stopped noticing I was paying. If your use of ChatGPT is personal and low-stakes, turn it on, enjoy it, and skip the rest of this article.
The reason I'm writing anyway is that “personal and low-stakes” describes almost none of how ChatGPT gets used inside a working business. And the upgrade that makes it better at remembering your preferences is the exact same upgrade that makes it better at remembering things you might not have meant to leave with it.
The bit that changes your data hygiene
Here's the shift in one sentence: the clean slate is gone. When memory was off, pasting a customer list or a supplier's pricing into ChatGPT to “just reformat this quickly” was a contained act — the information went in, you got your answer, the conversation was an island. With dreaming on, that same paste can become part of a synthesised, persistent picture the model carries forward and actively maintains. You've moved from a series of disconnected errands to a relationship with an assistant that takes notes and tidies them up on its own.
None of that is sinister. But it changes the question you should ask before you paste. It used to be “is this one conversation okay?” It's now “am I comfortable with this becoming something the tool remembers about my business?” Those are different questions, and most teams are still answering the old one out of habit.
The shadow problem most owners haven't clocked
The uncomfortable truth is your team is already doing this, with or without a policy. Surveys this year put small-business AI adoption around 82%, with the average firm using a handful of AI tools day to day. That isn't a future risk to plan for; it's a present reality to get tidy about. Someone in your business has pasted something into ChatGPT this month that you'd rather hadn't become permanent — a customer's details or a half-finished contract. They weren't being reckless. They were being efficient, on a tool that until last week forgot everything anyway.
I saw a gentle version of this with a client a couple of months back. Their team had been using a single shared ChatGPT login for everything — drafting marketing copy, yes, but also tidying up the occasional client spreadsheet. Nobody had done anything wrong, and no harm came of it. But once you picture that shared account quietly building a persistent memory across all of it, the fix becomes obvious: the marketing brainstorming and the client-data tidying should never have shared a brain in the first place.
The one rule I gave every client this week
I didn't hand anyone a fifteen-page policy. I gave them one rule with three small parts, because a rule people remember beats a document people file.
- Separate your “thinking” use from your “customer data” use. Use the assistant freely for drafting and planning — the stuff where memory is a pure win. For anything involving real customer, financial or staff data, either use a workspace where memory and training are off by default, or don't paste the raw data at all and work with anonymised or dummy versions. One brain for ideas, a different posture for data.
- Know what it remembers — actually look. Open the new memory summary page. Read what it thinks it knows. Delete anything that's drifted in that shouldn't be there. This takes five minutes and it's the first time most people see, in plain English, the picture the tool has quietly built. Do it once yourself before you ask your team to.
- Write the rule down in one line. Something like: “ChatGPT is for drafting and thinking. Real customer or financial data goes in our approved tools, not the chat.” That single sentence, said out loud in a team meeting, does more than any policy PDF nobody reads.
What I'd do this week
Three jobs, none of them heavy.
- Check whether you're on a shared login, and stop if you are. Shared accounts and persistent memory are a bad combination — you get one synthesised picture built from everybody's mixed-purpose use. Give people their own access, even on the free tier.
- Review the memory page on the accounts that matter. Yours first, then anyone whose role means they handle customer or financial information. Treat anything that looks like real client data sitting in there as a prompt to tighten the habit, not a disaster.
- Say the one-line rule out loud. Not an email people skim. A sentence in a meeting that everyone hears, so the next time someone's about to paste a client list to save two minutes, the rule is already in their head.
This isn't a reason to be frightened of ChatGPT or to ban it — that would cost you far more than it saves, and your team would route around it anyway. Dreaming is a good upgrade, and the transparency page is exactly the kind of thing I want to see more of from these companies. The tool grew a memory. The sensible response to anything growing a memory is to be a little more deliberate about what you tell it. The rule is one sentence. The memory review is five minutes. Do both, and a better ChatGPT becomes an asset. Skip them, and the same upgrade quietly becomes a liability.