How ChatGPT's 'Dreaming' Memory Works (and What to Turn Off)

ChatGPT's "Dreaming" memory is a background process that reads across all of your past conversations and quietly rewrites a living profile of who you are — your preferences, projects, and habits — without you saving anything. Launched as Dreaming V3 on June 4, 2026, to Plus and Pro users in the US, it replaces the old manual "saved memories" list with something closer to how a person forms long-term impressions: it synthesizes, weighs, and updates on its own. That's why it suddenly feels like ChatGPT knows you — and why it's worth understanding exactly what it stores.

Here's the part that surprises people: the model itself remembers nothing. Every time you hit send, ChatGPT starts from a blank slate. What feels like memory is a separate system bolted on top — and Dreaming V3 is the most aggressive version of it yet.

Table of Contents

1. The Core Trick: The Model Is Stateless
2. What Changed With "Dreaming V3"
3. How "Dreaming" Actually Works, Step by Step
4. Why It Feels Uncanny — and the Privacy Catch
5. What You Can Control (and Turn Off)
6. Frequently Asked Questions

---

The Core Trick: The Model Is Stateless

A large language model is stateless — it has no memory of you between messages. The only reason ChatGPT can refer back to something you said earlier in a conversation is that the entire chat so far is silently pasted in front of your new message every single time. That bundle of text is the context window, and when it fills up, the oldest parts fall off the edge. (We go deeper on this in why AI forgets what you talked about and context management.)

So "memory" across different chats was never the model — it was always a notebook. The question is who writes in the notebook, and how much.

• Old way (saved memories): ChatGPT wrote down a short list of facts you explicitly told it to remember, or that it flagged as important. You could read the list, edit it, delete lines. Transparent, but shallow.
• New way (Dreaming V3): a background process reads across years of your chats and continuously rewrites a rich profile of you — no prompting required.

What Changed With "Dreaming V3"

OpenAI's Dreaming V3 update, which began rolling out June 4, 2026, is a rethink of how the chatbot stores, weighs, and uses what it learns about you. Three things are genuinely new:

OpenAI reports the new system reached 82.8% task success on its factual-recall evaluation, and exposes three user-facing controls: a readable summary page of what it has synthesized about you, a way to add or correct details, and settings for which topics it should raise and when. (OpenAI rollout coverage)

How "Dreaming" Actually Works, Step by Step

The name is a metaphor for what the system does when you're not looking — the same way sleep is thought to consolidate human memory. Here's the loop:

1. Collect. Your conversations accumulate as raw transcripts tied to your account.
2. Synthesize ("dream"). On a schedule, a background model reads across those transcripts and writes a summary of you — not the raw chats, but distilled conclusions: "prefers concise answers," "is building a SaaS," "lives in a Pacific time zone." This is the step that's new and powerful.
3. Weigh and update. New conclusions can overwrite or sharpen old ones, so the profile stays current instead of piling up stale facts. Recent, repeated signals count more.
4. Inject. When you start a new chat, the relevant slice of that profile is quietly added to your context window — the notebook gets read back in — so the model answers as if it remembers you.

The model is still stateless at every step. The intelligence is in the librarian that decides what to write down and what to hand back.

Why It Feels Uncanny — and the Privacy Catch

Two things make Dreaming V3 feel different from a settings toggle.

It infers things you never stated. Because it synthesizes rather than transcribes, the profile can contain conclusions you didn't volunteer — drawn from patterns across many chats. That's what produces the "how did it know that?" moment.

The audit trail is thinner. With the old list, you could see every line. With a continuously re-synthesized profile, what you read on the summary page is itself a summary — independent reporting has noted the trade-off between a self-updating memory and a fully inspectable one. (TechTimes analysis) Privacy researchers have separately flagged that most memory entries in systems like this are created by the system unilaterally, not by the user.

None of this is inherently sinister — it's the same mechanism that makes the feature useful. But "useful" and "knows more about me than I told it" are the same coin.

What You Can Control (and Turn Off)

You are not stuck with it. In ChatGPT's settings under Personalization → Memory, you can:

• Read the summary page to see what it has synthesized about you, and correct or delete entries.
• Turn memory off entirely — ChatGPT then treats each chat as a clean slate (the stateless default).
• Use Temporary Chat for anything you don't want feeding the profile — these aren't "dreamed" into memory.
• Set topic controls for subjects you don't want it raising on its own.

A good habit: read your summary page once a month the way you'd check a credit report. It's the clearest window into what the AI thinks it knows.

Frequently Asked Questions

What is ChatGPT's "Dreaming" memory?

It's the background memory system OpenAI launched as Dreaming V3 on June 4, 2026. Instead of you manually saving facts, a background process reads across your past conversations and continuously writes and updates a synthesized profile of you, which is fed into new chats so ChatGPT can personalize its answers.

Does ChatGPT actually remember conversations now?

Functionally yes, but not in the model itself. The AI model is stateless — it forgets everything between messages. The "memory" lives in a separate system that stores a profile of you and re-injects the relevant parts into each new conversation's context window.

Why does ChatGPT know things I never told it?

Dreaming V3 synthesizes conclusions across many chats rather than just storing facts you stated. So it can infer preferences and details from patterns — which is why it sometimes surfaces things you don't remember telling it directly.

How do I turn off ChatGPT's memory?

Go to Settings → Personalization → Memory. You can view and edit what it has stored, turn memory off completely (each chat becomes a clean slate), or use Temporary Chat for sessions you don't want remembered.

Is ChatGPT's Dreaming memory a privacy risk?

It depends on your comfort level. The benefit — relevant, personalized answers — comes from the same mechanism that builds a detailed, partly inferred profile of you with a less complete audit trail than the old saved-memories list. Reviewing your memory summary page periodically and using Temporary Chat for sensitive topics are the practical safeguards.