OpenAI Launches Dreaming V3 Memory Synthesis System
OpenAI officially released Dreaming V3, a more powerful and scalable memory synthesis system for ChatGPT.
This system significantly enhances memory freshness, continuity, and relevance, addressing issues of outdated, accuracy, and scalability of memories accumulated by hundreds of millions of users over the years.
Dreaming V3 is now available to Plus and Pro users in the U.S., with plans to expand to more countries and Free and Go users in the coming weeks.
ChatGPT will launch the "Save Memory" feature in April 2024, with the first version of Dreaming introduced in April 2025. This V3 version is an upgrade based on the previous architecture.
The new system shows substantial improvements in factual recall, preference adherence, and long-term accuracy: factual recall success rate increased from 41.5% to 82.8%, preference adherence from 31.4% to 71.3%, and long-term accuracy from 9.4% to 75.1%.
OpenAI has optimized the computational power required for Free users by about five times, supporting larger memory capacity and coverage for free users.
Source: Public Information
ABAB AI Insight
OpenAI has continued to invest since launching the memory feature in 2024. The iteration of the Dreaming series follows its strategic path of transforming ChatGPT from a single conversation tool into a long-term personalized companion, with Sam Altman's team focusing on solving the challenge of maintaining long-term context.
From a capital perspective, OpenAI's significant reduction in memory computational costs (5x optimization) allows for rapid deployment to Free users, thereby expanding the monthly active user base and enhancing paid conversion rates, while providing Plus/Pro users with larger memory capacity to strengthen subscription stickiness and ARPU.
Similar to the evolution of knowledge graphs in Notion and Roam Research, ChatGPT is transitioning from a short-term conversational AI to a personal operating system with long-term memory and preference learning, directly competing with memory systems from Anthropic and Google.
Essentially, this represents a technological replacement: AI memory systems replace manual recording and context management by humans, with mechanisms that synthesize backend organization instead of simple storage, achieving reliable recall across years through multidimensional optimization, allowing large models to genuinely possess the capability of "continuously learning users," accelerating the shift from tools to intelligent partners.
ABAB News · Cognitive Law
Memory is not storage, but the ability to synthesize continuously.
When computational costs decrease by 5 times, free users become a strategic weapon.
When AI truly remembers you, the era of single conversations ends.