OpenAI launches Dreaming V3, ChatGPT now remembers on its own
OpenAI has started rolling out a new ChatGPT memory that quietly synthesizes and updates what matters from your chats, while using five times less compute.
- OpenAI began rolling out Dreaming V3 on June 4 for Plus and Pro users in the United States.
- Memory now works on its own in the background after a chat and updates older entries.
- Compute use has dropped fivefold, which opens the door to free users.
- By OpenAI's internal measurements, factual recall jumped to 82.8 percent.
Have you ever opened ChatGPT and wished it would just remember what you already told it? OpenAI says it has now solved exactly that. On June 4 the company started rolling out a new memory system called Dreaming V3, first for Plus and Pro users in the United States.
The difference from the old setup is that you no longer have to tell the assistant what to keep. Dreaming kicks in as a background process once a conversation ends, reviews what was said, and creates and revises entries on its own. Free users, Go users and other countries, the company says, will follow after this first wave.
How Dreaming actually remembers
The easiest way to explain it is the example the company itself gives. If you mention today that you are going to Singapore in July, the assistant notes it down. Once July passes, Dreaming rewrites that entry into a more accurate version, for instance that you were in Singapore in July 2026. The context stays current instead of turning into a pile of stale notes.
OpenAI describes how the system works:
Dreaming triggers a background process after a conversation ends, synthesizing what matters , your preferences, constraints, ongoing projects and time-sensitive context, automatically.
The numbers the company published back up that story. By its internal measurements:
- factual recall rose to 82.8 percent from 67.9
- tracking of your preferences jumped to 71.3 percent from 31.4
- data freshness climbed to 75.1 percent from 52
These figures are internal and have not been confirmed by independent testers, so read them as a vendor claim rather than a verdict. Still, the direction is clear, the assistant should get things wrong less often and forget less.
Why does five times less compute matter?
This is the real reason the news is more than just another upgrade. The old memory was expensive to run, which is why it mostly came with paid plans. OpenAI says Dreaming uses five times less compute for the same job.
When something costs five times less, it is far easier to give away. That is exactly why the company is signalling the feature will reach free users too. For the average person who uses ChatGPT for work, cooking or studying, it means an assistant that remembers more will no longer be a subscriber-only perk.
One question many are already asking remains. How comfortable are you with an assistant deciding, on its own and without your sign-off, what in your life is worth remembering?
