OpenAI recently announced a comprehensive upgrade to ChatGPT’s “Memory” feature, launching a memory synthesis system based on its next-generation Dreaming architecture. This update aims to completely resolve issues in the previous generation of Memory functions that emerged during long-term usage, such as outdated information, declining accuracy, and bottlenecks in cross-user scalability. The new system continuously integrates multi-turn dialogue context through background automated processes, enabling dynamic updates and automatic “freshness preservation” of memories, effectively eliminating the pain point where contexts become obsolete over time. Users can view, edit memory content, or set interaction rules at any time via the memory summary page. Official evaluation data shows that compared to the first-generation version launched earlier this year 2024, the new version has achieved breakthrough improvements across several metrics: fact recall rate increased to 82.8%, preference adherence reached 71.3%, and timeliness correction capability improved to 75.1%. In terms of commercialization, OpenAI significantly reduced inference costs by approximately five times through technical optimizations applied to the Dreaming architecture, making it viable for scaling to massive user bases for the first time. The feature has been prioritized for rollout to Plus and Pro subscribers in the United States, with plans to gradually extend coverage to global users and Free and Go free-tier accounts within the coming weeks. According to OpenAI, this move will unify the underlying memory infrastructure for all full-capacity users, laying the foundation for personalized AI services over the long term.