OpenAI Revolutionizes ChatGPT with Dreaming V3 and Enhanced Memory

Two years after the introduction of memory in ChatGPT, OpenAI announces the launch of Dreaming V3. This new architecture, introduced in April 2025 to complement manually recorded memories, now becomes the sole foundation of the assistant's memory. It aims to address the shortcomings identified among its hundreds of millions of users, particularly obsolescence, inaccuracy of memories, and long-term management difficulties.

Dreaming V3 enhances context continuity by allowing ChatGPT to retain useful information from one conversation to the next. Users' personal preferences, such as response style or work habits, are now better integrated and applied consistently. Additionally, memories evolve automatically, preventing information like dates from becoming static.

This technical overhaul also reduces the computational power required by a factor of five, making the deployment of Dreaming V3 viable for free accounts. A memory summary page has been introduced, providing users with an overview of the information retained by ChatGPT, with the option to modify or delete it, and to specify topics the assistant should address or avoid. This page replaces the old list of manually recorded memories.

The deployment of Dreaming V3 begins today for Plus and Pro subscribers in the United States, with doubled memory storage capacity. A gradual rollout to other countries and free accounts is planned in the coming weeks.