Dreaming : OpenAI recasts ChatGPT's memory

OpenAI updates ChatGPT with Dreaming, a memory background process boosting factual recall to 82.8% and cutting compute 5x for Plus, Pro, and Free users.

ChatGPT's memory architecture is changing. Dreaming, the background process that automatically synthesizes memories from conversation history without waiting for an explicit memory request, is transitioning to a significantly more capable and compute-efficient version. OpenAI is thus addressing the three weaknesses it observes across hundreds of millions of users: obsolescence, accuracy, and scalability.

Where 2024's saved memories functioned like static notes, destined to become false or irrelevant, Dreaming updates the state of memory over time. A trip planned for July becomes, once the month has passed, a trip taken, and recommendations readjust to the return location and time zone. A summary page also allows users to consult what the model retains, correct details, or specify which topics to address and when.

Internal evaluations quantify the leap: factual recall increases from 41.5% in 2024 to 82.8% today, adherence to preferences from 31.4% to 71.3%, and temporal accuracy from 9.4% to 75.1%. An approximately 5x reduction in necessary compute also opens the system to free accounts and increases the memory capacity for Plus and Pro subscribers. Deployment begins with Plus and Pro users in the United States, before other countries and Free and Go offerings in the coming weeks.