A New Era of Contextual Understanding
OpenAI has launched a major architectural upgrade to ChatGPT's memory system, introducing a new "dreaming" architecture that promises to revolutionize how the chatbot retains and utilizes information across conversations. This new system, known as Dreaming V3, is designed to make ChatGPT more capable, current, and scalable, moving beyond its previous reliance on explicit user instructions to remember details. The core goal of this upgrade is to tackle long-standing issues of outdated and insufficient accuracy in ChatGPT's memory, significantly enhancing its ability to provide personalized responses over extended periods.
The initial memory feature, launched in April 2024, functioned more like a "saved notes" system, requiring users to explicitly tell ChatGPT what to remember. This approach proved static and often led to memories becoming stale or irrelevant over time. In April 2025, OpenAI introduced the first version of "dreaming," a background process that allowed ChatGPT to reference chat context outside of explicitly saved memories. While an improvement, this initial dreaming system was not robust enough to serve as a standalone memory foundation. The latest 2026 update, built on top of this dreaming concept, aims to make it the central, more capable, and compute-efficient memory architecture.
Dreaming V3: Smarter, Fresher, More Relevant
The new Dreaming V3 system allows ChatGPT to learn from many conversations and synthesize its memory state in the background, ensuring that conversations always have the freshest and most relevant context. This means ChatGPT can now automatically curate memories from chat history without relying on explicit requests to remember something. OpenAI emphasizes three key areas of improvement with this new system: carrying forward useful context, following user preferences and constraints, and staying current with the passage of time. For instance, if a user mentions a trip to Singapore in July, the dreaming system will automatically update that memory to reflect when the trip is upcoming, happening, or already over, providing tailored recommendations based on the current context.
Internal evaluations by OpenAI demonstrate the significant strides made by Dreaming V3. Factual recall task success has improved from 67.9% in 2025 to 82.8% in 2026. Preference adherence has also seen a substantial jump, rising from 55.3% to 71.3%. Furthermore, the system's ability to stay current over time has reached 75.1%, a notable increase from 52.2% in 2025. These improvements collectively contribute to a more natural and personalized interaction with the AI, making it feel more like a genuine assistant that understands user needs and adapts over time.
Expanded Access and User Control
One of the most impactful aspects of this update is the expansion of the advanced memory features to ChatGPT's free tier users for the first time. This was made possible by significant efficiency improvements, with OpenAI engineers successfully reducing the compute resources required to run the background dreaming system by approximately five times. This optimization makes it economically viable to distribute these advanced memory capabilities to hundreds of millions of free accounts. While Plus and Pro users in the US are receiving the upgrade immediately, with their memory storage capacity doubling, the rollout to additional countries and Free and Go users is expected in the coming weeks.
Users will also have greater control over ChatGPT's memory. The new system includes a "memory summary" page where users can review what ChatGPT knows about them. From this summary, individuals can add, update, or correct information, and even provide instructions on when specific topics should be brought up. This level of transparency and control allows users to manage their personalized AI experience effectively. For conversations where memory is not desired, users can opt for a "Temporary Chat" which will not use or update memory.