OpenAI Enhances ChatGPT Memory with Advanced "Dreaming" Architecture

OpenAI has announced a major enhancement to ChatGPT's memory capabilities, deploying an upgraded "Dreaming" architecture that promises a more personalized and context-aware interaction for users. This new system is designed to significantly improve how the AI remembers details from past conversations, user preferences, and evolving information over time. The rollout of this advanced memory feature begins with ChatGPT Plus and Pro subscribers in the United States, with plans to extend availability to Free and Go users, as well as other international regions, in the upcoming weeks. This marks the third significant iteration of ChatGPT's memory functionality since its initial introduction in April 2024. The "Dreaming" architecture addresses previous limitations where users often had to reiterate information or preferences in new chat sessions. With this update, ChatGPT will automatically synthesize information from various interactions, creating a "memory summary" that users can review and edit. This allows for a more seamless experience, as the AI can proactively apply learned preferences and facts without explicit prompting. For instance, if a user discusses their camera setup, future inquiries about compatible gear will yield tailored recommendations. OpenAI reports substantial improvements in key performance metrics with the new Dreaming V3. Factual recall, the ability to remember specific user-shared information, has reportedly increased to 82.8%, a notable jump from 41.5% in 2024 and 67.9% in 2025. Similarly, preference adherence, or the system's capacity to follow user-defined constraints and behaviors, has improved to 71.3% from 31.4% in 2024. These advancements aim to make ChatGPT more efficient and less prone to providing stale or contradictory information. Users will also gain access to a "memory sources" feature, allowing them to see and manage the information ChatGPT utilizes for personalization.