On June 4, OpenAI quietly began rolling out the biggest change to ChatGPT’s memory since the feature existed — and it did not announce it like a privacy decision, because OpenAI does not want you to think of it as one. It’s called Dreaming V3, and it does something the old memory never could: it reads across years of your past conversations on its own, builds a running profile of who you are, and updates that profile while you’re not even using the app. You don’t tell it what to remember anymore. It decides.
And here is the part buried under the press-friendly stats: deleting your conversations no longer deletes what ChatGPT learned from them. The old memory was a notepad you controlled. Dreaming V3 is a private investigator who never clocks out — and won’t show you the full file.
OpenAI Just Quietly Redefined What “Delete” Means
Until June 4, ChatGPT memory was two simple layers: an explicit list of facts you asked it to save, and a weaker background process called Dreaming V0 that could glance at your chat history but couldn’t stand on its own. Dreaming V3 collapses both into a single asynchronous engine that synthesizes memory from many conversations at once and rewrites entries as your life changes. OpenAI’s own example: a note reading “you’re going to Singapore in July” silently rewrites itself to “you went to Singapore in July 2026” once the trip passes. No prompt. No permission. The system just knows.
That temporal awareness is the headline feature. The consequence nobody is putting on a slide is structural: this synthesized memory is not stored inside your conversation log. It lives in a separate data layer and gets injected into the system prompt at the start of every new chat. So when you delete a conversation to clean up after a sensitive exchange — a health scare, a financial mess, a fight with your partner — the source vanishes, but the inference ChatGPT already drew from it stays. As Tech Times reported on June 5, removing a chat leaves the derived memory intact; to actually erase a detail you have to delete both the memory entry and the original conversation. Even then, OpenAI’s own policy says logs of deleted memories can be retained for up to 30 days.
Read that again. The thing you did to protect yourself — hitting delete — now does roughly half of what you think it does.
The Numbers OpenAI Wants You to See — and the One It Doesn’t
OpenAI is selling this on performance, and the numbers are genuinely impressive. Factual recall, the company says, jumped from 41.5% in the 2024 notepad era to 82.8% with Dreaming V3, with preference adherence at 71.3% and time-sensitive accuracy at 75.1%. A roughly 5x cut in compute cost is what makes it cheap enough to push to free users worldwide; Plus and Pro subscribers get 2x the memory capacity on top.
Every one of those figures is OpenAI’s own, measured on OpenAI’s internal benchmarks, confirmed by no independent auditor. That’s worth saying plainly because the company is asking for a remarkable amount of trust: let an unaudited system build a persistent psychological profile of you, and trust its self-graded scorecard that it’s doing so accurately.
The number OpenAI doesn’t advertise is the one that matters most. The new Memory Summary Page lets you view and correct what the system thinks it knows — but OpenAI itself admits the summary does not necessarily show everything ChatGPT remembers. And clicking “Don’t mention this again” only suppresses a detail from surfacing; the underlying entry stays in the data layer unless you explicitly hunt it down and delete it. You are being handed a dashboard that is, by the vendor’s own admission, incomplete. A 2026 study presented at the ACM CHI Conference named the trap exactly: a “personalization-convenience paradox,” where the feature people value most is also the one they cannot fully audit or constrain.
“Dreaming” Is the Most Honest Name OpenAI Has Ever Picked
Give OpenAI credit for the branding, because it told on itself. The name nods to how the human brain consolidates the day’s events during sleep — and that’s precisely the unsettling part. The system processes you in the background, drawing connections you never explicitly drew, when you have no idea it’s running. Users are already coining a term for the result: “context bleed,” where a health detail from one chat quietly reshapes a diet suggestion in another, or a money worry you resolved months ago keeps tinting the answers you get today.
To be fair — and this is the line that keeps it honest — context bleed is not a data leak between accounts. Your memories stay scoped to you; no evidence has emerged of one user’s profile reaching another. The system is, in most cases, working exactly as designed. That’s the problem. The friction isn’t a bug. It’s the gap between OpenAI’s judgment of what’s “relevant” to resurface and your own expectation that what you said in one room stays in that room.
India’s Users Are Walking Into This Without a Net
Here’s where it gets uncomfortable for the country with one of ChatGPT’s largest user bases. In Europe, Dreaming V3 walks straight into a wall: under GDPR, a system that builds persistent behavioral profiles is doing “profiling,” which triggers consent duties and a hard right to erasure. The EU AI Act’s transparency rules for chatbots take effect on August 2, 2026 — under two months after this rollout — forcing OpenAI to meet real disclosure standards almost immediately. Italy’s regulator already fined the company €15 million in 2024 and has shown it will use the tools it has.
India has no such backstop arriving on a calendar. The Digital Personal Data Protection Act exists on paper, but its rules are still being operationalized and its enforcement teeth remain largely untested against exactly this kind of cross-context AI profiling. The United States, for its part, has no federal privacy law governing chatbot memory at all. So the protection you get from Dreaming V3 depends almost entirely on where you happen to live — and hundreds of millions of Indian users are getting the same silently-profiling system with the weakest legal floor underneath it. Who gets hurt? The users in the jurisdictions where “we’ll sort out the rules later” is the actual policy.
The Verdict: A Genuinely Great Feature With a Consent Problem at Its Core
Don’t mistake this for a panic piece. Dreaming V3 is, on raw capability, the most useful memory upgrade ChatGPT has ever shipped — an assistant that actually remembers your context is worth a lot, and most people will love it. But OpenAI made a deliberate choice to flip the default from “remember what I ask” to “remember everything and let me opt out of the parts I can find.” That is a different product with a different relationship to your privacy, and shipping it as a quiet quality-of-life update rather than the consent decision it actually is tells you how OpenAI wants you to feel about it: grateful, not cautious.
If you use ChatGPT for anything personal, do two things this week. Open Settings and actually read your Memory Summary — then assume it’s incomplete. And put genuinely sensitive conversations in Temporary Chat, the one mode that doesn’t read or write memory at all. The controls exist. OpenAI just built a system that bets, correctly, that almost no one will use them.