Imagine: you once told ChatGPT about dietary restrictions or a planned trip — and never returned to that topic. The new system will remember it on its own, apply it in future conversations, and even recognize if circumstances have changed. This is exactly what OpenAI promises with the Dreaming V3 update.
From Note Lists to Background Learning
The first versions of ChatGPT's memory worked simply: the model kept a list of facts that the user or the model itself added to it. In April 2025, OpenAI launched the first generation of "dreaming" — a background process that automatically synthesizes memories from chat history without explicit user commands. Now the company has rolled out Dreaming V3 — a substantially more efficient version of the same idea.
"Dreaming is a method by which ChatGPT automatically forms memories in the background, drawing on chat history"
— OpenAI
According to the company's internal assessments, the accuracy of fact reproduction increased from 67.9% to 82.8%, adherence to preferences — from 55.3% to 71.3%, and the relevance of information over time — from 52.2% to 75.1%. Separately, OpenAI reports a fivefold reduction in computational costs for maintaining dreaming for free users — this is what allowed the feature to be unlocked for the free tier.
Memory Summary: Transparency with a Caveat
Along with the architecture came the Memory Summary page — a consolidated overview of what ChatGPT knows about you. Here you can view, edit, and delete individual memories. But there's a detail that's easy to miss: deleting a conversation does not delete the memories extracted from it. They must be erased separately — manually, through settings.
This is not news for those who read the documentation carefully. But in the context of Dreaming V3, which now more actively scans past chats and synthesizes new records from them, the scope of a potentially "invisible" user profile grows.
- Free plan: memory is now available with the new architecture
- Plus / Pro: full access to Memory Summary with editing
- Enterprise / Edu: the Reference Chat History feature is currently unavailable
- Temporary Chat: conversations are not saved and not used for training — the only mode with clear isolation
Context That OpenAI Doesn't Advertise
Parallel to the technical update, litigation in The New York Times v. OpenAI case was ongoing. In May 2025, a federal judge ordered the company to preserve all ChatGPT logs — including those users had already deleted. The preservation order was lifted only at the end of September 2025, after which OpenAI resumed its usual deletion practices. This episode clearly shows: control over "your" data in a chatbot is always partial.
OpenAI emphasizes that users can disable memory entirely or use Temporary Chat for sensitive conversations. But the choice between the convenience of personalization and data minimization — is a very real compromise, not a marketing formality.
If Dreaming V3 truly increases memory accuracy to 82% and continues to scale — the question is not whether ChatGPT will become more useful, but whether OpenAI will develop a mechanism that allows users to see the full profile compiled from synthesized conversations — not just the fragments the system has decided to show.
