OpenAI Revamps ChatGPT’s Memory System, Cutting Computation Costs by 80%

OpenAI, a globally recognized leader in artificial intelligence, has recently unveiled a significant upgrade to its core interaction capabilities. To effectively tackle the long-standing challenges of outdated memory and information bias in extended AI conversations, the company has thoroughly restructured the underlying architecture of ChatGPT's memory system. This improvement not only enhances the accuracy of the information provided but also slashes computational costs to just one-fifth of their previous level.

Examining the development of ChatGPT's memory system reveals two key technological advancements. The initial version was launched in April 2024, featuring a basic “saved memories” function that depended heavily on user prompts. Users had to explicitly instruct the system with statements such as “Please remember something,” which allowed it to store only limited pieces of information in the background. By April 2025, OpenAI introduced a “Dreaming” mechanism for testing purposes within this memory system. This technology enabled the system to automatically review, filter, and organize users’ chat history through background processes, reducing reliance on specific prompts and significantly boosting personalization levels. However, the technical framework at that stage was not yet capable of supporting a fully functional memory system on its own.

The core innovation behind this upgrade is the transformation of these technologies into a more mature and efficient underlying architecture known as Dreaming V3. This new system is designed around three fundamental principles: maintaining relevant context, aligning with user preferences and constraints, and enabling continuous updates over time. Not only does this make the AI’s long-term memory more capable, but it also reduces the strain on system scalability when handling hundreds of millions of users across extended usage periods.

From a practical interaction standpoint, this underlying improvement brings noticeable improvements to the user interface. The new system automatically compiles important details into concise “summaries” displayed on a dedicated page, allowing users to quickly review what ChatGPT has learned about them—much like examining their personal file. Users can also directly ask the model questions in a conversational manner, similar to talking to an old friend, to learn more about specific memories or to adjust, modify, or remove outdated personal information at any time. This gives them precise control over which topics should be included in particular conversations.

As of now, this updated memory system has been fully deployed for Plus and Pro paid users in the United States, doubling their core memory capacity. Most notably, thanks to engineering breakthroughs that reduced the computational requirements of the Dreaming service by 80%, OpenAI has announced that it plans to roll out this feature to its global free user base in the near future, as long as service quality remains intact. This move will help accelerate the advent of fully personalized, high-precision AI experiences for all users.