ChatGPT Now Remembers Everything, Using Much Less: Memory Function Updated

OpenAI has begun rolling out a major update for ChatGPT's memory management, aimed at addressing scalability, obsolescence, and accuracy issues of long-term accumulated information. Developed on top of the background infrastructure called "dreaming" (initially launched in April 2025 to automatically synthesize useful data without requiring manual saving), the new architecture reduces the computational resources needed for processing historical contexts by five times.

This drastic reduction in computing costs finally allows the functionality to be extended to free accounts as well. Until now, dynamic memory processing was exclusive to Plus and Pro subscribers. Thanks to infrastructure optimizations, the dreaming-based technology will reach users of the free version of the virtual assistant in the coming weeks. For paying users in the United States, however, the new architecture goes live today, accompanied by an immediate doubling of personal data storage capacity.

The "Dreaming" Architecture Lowers Costs and Scales to Millions of Users

The first memory models distributed in February 2024 required manual input of explicit instructions to save writing preferences or planning details, a cumbersome system that quickly led to information obsolescence. The shift to automatic extraction via dreaming resolved usability issues, but it is only with the current architecture that OpenAI claims to have achieved the necessary efficiency on a global scale.

"We’ve been researching new ways for ChatGPT memory to carry context across conversations and keep it useful over time. Today, that work is rolling out as a more capable memory system in ChatGPT."
— OpenAI (@OpenAI) June 4, 2026

The technical breakthrough is supported by internal metrics provided by OpenAI, which highlight significant performance improvements compared to data recorded in 2025. Factual recall has risen from 67.9% to 82.8%, while the model's ability to adhere to user preferences has increased from 55.3% to 71.3%. There is also a notable increase in the temporal accuracy of the information, reducing past errors stemming from the use of outdated details.

User interaction with this cognitive archive occurs through the new memory summary page. Within this dashboard, users can quickly review highlights from their profile, correct or update incorrect information, and set constraints on the topics the AI should recall and under what specific circumstances. To delve deeper or modify a specific topic, users simply need to initiate a normal chat session with the assistant.

The update aims to refine three key areas: the fluid retention of useful contexts, adherence to user-defined limits, and the maintenance of up-to-date information over time. After the initial release in the United States for Plus and Pro customers, the rollout will proceed globally to progressively include Free and Go accounts.