OpenAI has launched a new generation ChatGPT memory system called Dreaming to optimize novelty, coherence, and relevance.

BlockBeats News, June 5th, OpenAI announced the launch of a more powerful and scalable ChatGPT memory synthesis system Dreaming, aimed at enhancing the freshness, continuity, and relevance of memories, addressing the staleness, accuracy, and scalability issues that arise when applying memories across hundreds of millions of users and years of timespan. OpenAI stated that the memory feature will help ChatGPT learn user preferences, context, and constraints to enable future conversations to unfold based on shared context rather than starting from scratch. This update is now available to U.S. Plus and Pro users today and will be rolled out to more countries and Free and Go users in the coming weeks.

Introduced in April 2024, ChatGPT initially launched the "Save Memory" feature, allowing users to request ChatGPT to remember specific information. In April 2025, OpenAI introduced the first version of Dreaming, enabling ChatGPT to reference chat history through a background process, automatically organizing and synthesizing memory states. The newly released Dreaming V3 is a next-generation memory architecture built on this mechanism.

OpenAI stated that the new system has seen improvements in fact recall, preference adherence, and maintaining accuracy over time. Their evaluations show that the fact recall task success rate has increased from 41.5% in 2024 to 82.8% in 2026; preference adherence success rate increased from 31.4% to 71.3%; and maintaining accuracy over time success rate increased from 9.4% to 75.1%. OpenAI noted that recent improvements have reduced the computing power required for Free users to access Dreaming by approximately 5 times, enabling its rollout to Free users in the coming weeks and enhancing memory capacity for Plus and Pro users.