OpenAI's 'Dreaming' Gives ChatGPT a Memory That Updates Itself

OpenAI started rolling out Dreaming V3 on June 4, its biggest overhaul of ChatGPT's memory since the feature launched in 2024. The update lets ChatGPT build and refresh what it knows about you automatically — and for the first time, it's coming to free users too.

The rollout began with Plus and Pro subscribers in the United States and is expected to expand to Free and Go users in the coming weeks. If you use ChatGPT regularly, this is the change most likely to alter your day-to-day experience.

How ChatGPT's memory got here

ChatGPT's memory has come in three stages. In February 2024, OpenAI added saved memories — you had to explicitly tell the model "remember this," and it filed the fact away. In April 2025, the system started referencing your full chat history, not just the notes you flagged.

Dreaming V3 is the third stage, and it flips the model from passive to active. Instead of waiting for instructions, ChatGPT now runs a background process after your conversations end. It reviews what happened and decides on its own what's worth keeping.

Why this matters: the old approach treated every saved fact as permanent. That's the root of a familiar annoyance — an assistant that keeps acting on stale information long after it stopped being true.

What's actually new

OpenAI says Dreaming optimizes memory along three dimensions: freshness, continuity, and relevance. The headline capability is that memories now decay and update over time rather than sitting frozen.

The company's own example: if you tell ChatGPT "I'm going to Singapore in July," the old system kept that as a standing fact and might recommend Singapore restaurants forever. Dreaming revises it after the trip to "You went to Singapore in July 2026" — and stops pushing recommendations you no longer need.

The synthesis happens automatically, pulling signal from across many conversations without you asking. The point is to capture your preferences, constraints, and ongoing projects as they evolve, so the assistant feels less like a notepad and more like something that actually tracks your life.

You stay in control. Users can view, edit, or delete any stored memory, and temporary chats remain fully excluded — nothing from those sessions is stored or referenced later.

Why free users are the real story

The most consequential detail isn't the architecture — it's who gets it. OpenAI extended Dreaming to the free tier because it cut the compute cost of running the synthesis process by roughly 5x.

That's an economics story as much as a product one. Background memory synthesis is expensive to run at ChatGPT's scale, and the cost is what kept richer memory locked behind paid plans. Bringing it to free users signals OpenAI is confident it can serve the feature cheaply enough to make it the default everywhere.

Paid tiers still get a sweetener: Plus and Pro subscribers receive double the memory storage capacity.

Why this matters for OpenAI

This is a retention play, not a benchmark move. OpenAI's biggest competitive risk isn't that ChatGPT can't solve hard problems — rivals like Google's Gemini and Anthropic's Claude have closed much of the capability gap. The risk is that people drift away from an assistant that keeps asking them to repeat themselves.

A memory that quietly keeps up with your life raises switching costs. The more ChatGPT knows about your projects and preferences, the more starting over with a competitor costs you. That's a durable moat that has nothing to do with raw model performance.

The privacy questions

Memory systems that infer behavioral profiles invite scrutiny, and Dreaming is already drawing it. The core tension: a system that decides on its own what to remember is, by design, building a profile of you without explicit consent at each step.

A February 2026 arXiv study found that 96% of ChatGPT memories in a sample of roughly 2,050 entries from 80 users had been created unilaterally by the system, without the user prompting them. Dreaming makes that dynamic the norm rather than the exception.

This collides with regulation on the horizon. The EU AI Act's transparency rules take effect in August 2026, and systems that automatically build user profiles will need to explain what they're doing. Expect that to shape how Dreaming surfaces and discloses what it stores in Europe.

What's next

Watch the expansion timeline. The free-tier rollout is the part that puts automatic memory in front of hundreds of millions of casual users, many of whom won't realize their conversations are being synthesized in the background. How prominently OpenAI surfaces the memory controls will determine how much pushback it gets.

Also watch for competitive responses. Persistent, self-updating memory is becoming table stakes for consumer AI assistants, and Google and Anthropic will face pressure to match it.

Bottom line

ChatGPT now remembers you the way a good assistant would — keeping track of what's current and letting the rest fade. If you use ChatGPT, it's worth opening your memory settings to see what it has stored and prune anything you'd rather it forget. The convenience is real, but so is the fact that the system is now deciding what to keep on its own.