OpenAI Launches Dreaming V3 Memory System for ChatGPT

OpenAI is overhauling how ChatGPT remembers things. The company today began rolling out a new, more scalable memory architecture designed to address the staleness, correctness, and scalability challenges that emerge when memory is applied across hundreds of millions of users and years of conversations.



The upgrade builds on ChatGPT's existing "dreaming" mechanism, introduced as part of a memory update in April 2025. Dreaming uses a background process to learn from past conversations and synthesize ChatGPT's memory state. OpenAI says the new Dreaming V3 architecture is better at carrying forward useful context, following user preferences and constraints, and staying current as time passes. For example, if you previously told ChatGPT about dietary restrictions or an upcoming trip, the assistant should be better able to recall and apply that information in future chats while also recognizing when circumstances have changed.


A new memory summary page lets users review, add, or update what ChatGPT knows about them. OpenAI says the feature offers a more transparent way to manage memory, providing a high-level overview of stored information while making it easier to correct or expand specific details.



The update also brings a notable efficiency improvement. OpenAI says recent optimizations reduced the compute required to serve dreaming to free users by roughly 5x, making it practical to begin rolling the feature out to the free tier over the coming weeks. Plus and Pro users in the United States are getting the upgraded system starting today. OpenAI previously improved memory controls with the release of GPT-5.5 Instant, and today's update expands how ChatGPT carries context between conversations.

Free-tier users will gain access to dreaming-based memory, a capability that has until now been limited to subscribers. OpenAI says the rollout will expand to additional countries and to ChatGPT Go users in the weeks ahead.