OpenAI has launched Dreaming V3, a new memory synthesis system for ChatGPT that can automatically learn, organize, and refresh a user’s preferences and conversation context. The feature is now available to Plus and Pro users in the United States, and OpenAI said it plans to extend access to free-tier users in the coming weeks.
Memory moves beyond manual saving
OpenAI described Dreaming V3 as the latest step in ChatGPT’s memory development. In April 2024, the product relied on “Saved memories,” which required users to enter details manually. In 2025, “Dreaming V1” introduced background processing that could organize conversations automatically. With Dreaming V3, OpenAI says the system now runs on a new architecture with stronger computational efficiency and can operate as ChatGPT’s primary standalone memory system.
The release also includes a “Memory summary page,” where users can review, add, or edit what ChatGPT knows about them. They can also set triggers for specific topics. In OpenAI’s example, the assistant can remember that a user likes wildlife photography, prefers quiet dinners, and chooses hotels with strong air conditioning, then use that information to tailor travel suggestions.
Three tests for useful memory
OpenAI said the system evaluates memory quality using three core goals. The first is preserving useful context, so details from earlier conversations can still matter later. The second is following user preferences and constraints, making responses better aligned with personal habits. The third is keeping memory current over time, which is meant to reduce cases where stored information becomes outdated or inaccurate.
OpenAI illustrated this with a travel scenario: after a user returns home from Singapore, the updated system should stop suggesting Singapore restaurants at 5 a.m. and switch back to nearby local options. That example points to a common weakness in earlier memory tools. Information could be retained, but not always updated in line with real-world changes.
Compute cost for free users cut by about five times
On scalability, OpenAI said it has reduced the compute cost of offering the feature to free users by about 5x. That drop is a key part of making wider deployment possible. For now, Dreaming V3 is rolling out first to U.S. Plus and Pro subscribers, with more countries and users on the Free and Go plans expected to follow.
OpenAI’s framing makes clear that memory is no longer treated as a side feature inside ChatGPT. With Dreaming V3, it is being positioned much closer to the center of the product, and its performance at larger user scale will draw close attention.