OpenAI's Dreaming V3 Gives ChatGPT a Living Memory at 5x Less Compute
OpenAI's new Dreaming V3 architecture quietly synthesizes context from past chats, doubles memory capacity for Plus and Pro, and reaches Free users next.
Takeaways
- OpenAI rolled out Dreaming V3, a new background memory system for ChatGPT.
- Available to Plus and Pro users in the US today, with 2x memory capacity.
- Free and Go users get access in coming weeks after a 5x compute reduction.
- System automatically synthesizes context from chats without explicit "remember this" prompts.
- New memory summary page lets users review, edit, and steer remembered details.
- Memories now age with time, updating "trip in July" to "trip in July 2026" after it ends.
ChatGPT's memory just got a major architectural overhaul. OpenAI is rolling out a new system called Dreaming V3, which replaces the old approach of jotting down facts you explicitly asked it to remember with a background process that synthesizes context from your conversations and keeps that context fresh as time passes.
The goal is to tackle the staleness, correctness, and scalability challenges that show up when memory is applied to hundreds of millions of users and multi-year time horizons in ChatGPT. In practice, that means ChatGPT should stop forgetting things you mentioned offhand months ago, and stop confidently telling you about your trip to Singapore weeks after you got home.
Who gets it and when
The update is available to Plus and Pro users in the US today, and will roll out to additional countries and Free and Go users over the coming weeks. Plus and Pro users get 2x more memory as part of the rollout. To turn it on, update the ChatGPT app on iOS or Android. If you prefer the legacy saved memories experience, you can switch back in settings.
From sticky notes to a living profile
To understand why this matters, it helps to know how ChatGPT memory has evolved. Memory first launched in April 2024 as saved memories, which required explicit user instructions to store information. OpenAI introduced the first version of dreaming in April 2025, allowing ChatGPT to reference chat context beyond the saved memories list.
Saved memories worked, but only barely. They were closer to notes than a real understanding of a user. If you told ChatGPT you were traveling to Singapore in July, it could retain that fact. But the system was still dependent on strong cues, and those facts could grow stale.
Dreaming flips that model. It operates silently between interactions, allowing the model to synthesize insights from multiple conversations without manual prompting. You no longer need to say "remember this" for ChatGPT to notice that you shoot wildlife photography, live near Portola Valley, or hate crowded bars.
What good memory actually means
OpenAI evaluates the system against three concrete objectives, and each one corresponds to a real failure mode of the old approach:
- Carrying forward context. You tell ChatGPT something once and it shows up in later chats. In OpenAI's demo, asking about underwater photography gear yields a generic compatibility checklist without memory, but with memory it recommends a specific TTL trigger matched to the user's Sony A1 II, Nauticam NA-A1II housing, and Inon Z-330 strobes.
- Following preferences and constraints. A vegetarian gets vegetarian meal prep. A Singapore trip planner who likes wildlife photography and strong hotel AC gets a Mandai wildlife day, not a generic Merlion checklist.
- Staying current over time. The chatbot organically carries forward context, follows personal constraints, and automatically updates user preferences as time passes. The Singapore trip becomes a past trip once the dates pass, so a "what's open for takeout tonight" query routes back to your home city.
The compute breakthrough that unlocks Free
The reason Free users are getting access at all comes down to engineering efficiency. The biggest technical triumph behind this release is a massive optimization breakthrough. OpenAI engineers successfully reduced the compute resources required to run this background dreaming system by a staggering 5x. This makes it economically viable to distribute advanced memory features to hundreds of millions of free accounts.
That same efficiency gain is what funds the doubled memory capacity for paying tiers. A background synthesis process that runs continuously across hundreds of millions of accounts is expensive, and a 5x reduction is the difference between a Plus-only luxury and a default feature.
Visibility and control
One legitimate concern with a system that infers things about you in the background is opacity. OpenAI's answer is a memory summary page. The memories synthesized by dreaming are reviewable through a summary made visible in the memory summary page. From there, you can quickly glean the highlights of what ChatGPT knows about you, add or update information about yourself, and provide instructions on what topics ChatGPT should bring up and when. If you want to drill down into a particular area to learn more, just chat with the model.
There's a catch worth flagging. OpenAI says users can view a memory summary, make corrections, turn memory off, use Temporary Chat, and see some memory sources used to personalize responses. It also says the summary may not show everything ChatGPT remembers, because the system is based on a broader synthesis of context. If you want to fully scrub a piece of information, you may need to delete it from past chats, files, and connected apps too, not just the summary.
Why this is more than a feature update
The shift from saved memories to dreaming is a shift in what an AI assistant is for. A session-only assistant has to be reintroduced to your project, your stack, your constraints every time. A memory-rich assistant starts every conversation with that context already loaded.
For anyone using ChatGPT as an ongoing collaborator on code, writing, or research, the practical effect is that the assistant should stop asking what framework you use, stop suggesting libraries you've already rejected, and stop forgetting the architectural decisions made in last week's chat. Looking ahead, dreaming now provides a shared memory foundation for all users. This update represents OpenAI's most capable memory system yet. The deeper bet is that personalization, not raw model capability, becomes the moat.