Global AI leader OpenAI has officially announced a major upgrade to its core interaction experience. To completely address the two long-standing issues of "outdated memory" and "information bias" in AI long conversations, OpenAI has comprehensively renovated the underlying architecture of ChatGPT's memory system, significantly improving information accuracy while drastically reducing computational costs to one-fifth of the previous level.

Looking back at the evolution of ChatGPT's memory system, its foundational technology has undergone two core advancements. This feature was first launched in April 2024, with an early "saved memories" mechanism that heavily relied on user prompts, requiring humans to explicitly issue a command like "Please remember something," allowing the system to store only limited information fragments in the background. By April 2025, OpenAI first introduced a "Dreaming" mechanism for testing in the memory system. This technology could automatically review, filter, and organize users' chat history through background processes, no longer overly relying on rigid prompts, thus significantly enhancing personalization. However, the technical form at that time was not yet sufficient to support a complete memory system on its own.

image.png

The hard-core aspect of this upgrade lies in reengineering this technology into a more mature and efficient underlying architecture—Dreaming V3. The new system is deeply optimized around three golden standards: "continuing useful context, following user preferences and limitations, and dynamically updating over time." Not only does this make the AI's long-term memory smarter, but it also significantly reduces the system scalability pressure when facing hundreds of millions of users and multi-year usage cycles.

In terms of actual interaction experience, this underlying change brings highly humanized front-end changes. The new system automatically aggregates important details into high-density "summaries" displayed on a dedicated page. Users can quickly review what about themselves ChatGPT has learned, just like reviewing their personal file. Moreover, users can ask the model directly, as if chatting with an old friend, to inquire about the origins and details of specific memories or to add, modify, or manually delete outdated personal information at any time, finely configuring which topics should be mentioned in specific scenarios.

Currently, this new memory system has been fully rolled out to Plus and Pro paid users in the United States, with the core memory capacity doubling directly. Most excitingly, thanks to the engineering breakthroughs that successfully reduced the computing power consumption of the Dreaming service by 80%, OpenAI has clearly stated that, under the premise of ensuring service quality, this feature will also be officially rolled out to the global vast free (Free) user base in the near future, accelerating the arrival of the era of full-scale high-precision personalization for edge-side AI.