OpenAI recently announced a major architectural upgrade to ChatGPT’s memory system, which has significantly improved long-term context retention, user preference following, and cross-time accuracy. It also emphasized that this version is also more optimized in terms of computing power efficiency. The new system will first be available to ChatGPT Plus and Pro subscribers in the United States, and will gradually expand to more countries and Free and Go users in the coming weeks.
The so-called "memory" function means that ChatGPT can retain useful information in the conversation automatically or according to the user's instructions for a long time, including the user's personal preferences, ongoing projects, and various constraints, thereby avoiding having to introduce the background from scratch every time a new conversation is opened. With the help of this persistent information, the system can better "continue the context" in subsequent conversations and provide answers that are more in line with the user's habits and needs.
OpenAI will launch a "save memory" feature as early as February 2024, allowing users to explicitly ask ChatGPT to remember certain information, such as travel plans, writing style preferences, etc. However, this first-generation solution has obvious limitations: on the one hand, it is highly dependent on the user's active instructions to save information; on the other hand, if the information stored in the memory is not updated for a long time, it will easily become outdated or invalid.
By April 2025, OpenAI expanded its memory capabilities beyond “saved lists” and began allowing ChatGPT to automatically reference past chats as memory sources. The key infrastructure at this stage is a background process called "Dreaming", which automatically filters and organizes available information from historical conversations to form memory entries without the user being aware of it. With this mechanism, the system can continuously learn from natural conversations without the user having to explicitly say “remember this” every time.
The upgrade released this time further reconstructs the memory architecture based on the "Dreaming" mechanism, making it more "memorable, easy to use, long-lasting, and economical." OpenAI internal evaluation data shows that the new system has significant improvements over the 2025 version in fact recall, preference following, and accuracy over time. Specifically, the fact recall rate increased from 67.9% in 2025 to 82.8% in 2026, the compliance with user preferences increased from 55.3% to 71.3%, and the accuracy across time dimensions increased from 52.2% to 75.1%.
At the user experience level, an intuitive change brought about by this upgrade is the addition of the "Memory Overview Page". Users can centrally view various memory entries currently saved by ChatGPT on this page, and can directly update, correct them, and even give instructions on "which topics should be mentioned proactively under what circumstances." This design not only improves transparency, but also provides users with more fine-grained control, reducing the risk of memory obsolescence or bias accumulation during long-term use.
OpenAI said that this more powerful memory system will be first pushed to ChatGPT Plus and Pro users in the United States and will be online starting today. The company plans to gradually promote the system to more countries and regions in the next few weeks, and cover ChatGPT Free and Go users, so that a larger user group can experience conversational AI that better "understands themselves" in daily use.