OpenAI has announced an upgrade to ChatGPT's memory feature. Its memory technology "Dreaming," which retains conversational context and user preferences, has been updated to "V3," carrying past exchanges over more accurately while automatically rewriting memories as time passes. The rollout begins with Plus and Pro users in the United States and will gradually extend to other countries and to Free and Go users over the coming weeks.

The "Three Conditions" Good Memory Should Meet

ChatGPT's memory feature remembers a user's preferences, ongoing projects, and constraints so that each new conversation does not have to start from scratch. OpenAI lists three conditions that useful memory should satisfy: correctly recalling facts shared in past conversations, returning answers that respect preferences and constraints, and keeping memories from going stale over time.

For example, if you have mentioned that you are vegetarian, meal suggestions will naturally follow suit. A plan you described as "going to Singapore in July" is automatically rewritten to "went to Singapore in July 2026" once the trip is over. The heart of this update is the ability not just to remember, but to update the memory itself as circumstances change.

"Dreaming V3" Carries Memory Over More Accurately

The updated "Dreaming" is the internal technology that organizes the key points of a conversation and commits them to memory. With V3, past context and preferences are carried over with greater precision, and information that is no longer needed is updated and tidied up over time.

Earlier memory tended to blur the question of when a piece of information was shared, so finished plans and outdated preferences could linger. V3 addresses this weakness with a design that emphasizes keeping memory current.

Accuracy Gains Shown in Internal Evaluations

According to OpenAI's internal evaluation, Dreaming V3 posted large gains across all three conditions. The rate of correctly recalling facts shared in past conversations rose from 41.5 percent in 2024 to 82.8 percent. The rate of returning answers aligned with preferences and constraints improved from 31.4 percent to 71.3 percent, and the rate of updating memory in step with the passage of time climbed from 9.4 percent to 75.1 percent.

The metric for whether memory can be updated stands out in particular, jumping from 9.4 percent to 75.1 percent, roughly eightfold. That points to a major step forward in keeping memory current rather than hoarding outdated information.

A "Memory Summary Page" to Review What Is Stored

Alongside the upgrade, a new "memory summary page" lets users see at a glance what ChatGPT remembers about them. On this screen you can review the key points of what is stored, add or correct information, and specify topics or timing you would like ChatGPT to bring up. Because instructions can be given for each item, such as "edit this" or "stop bringing this up," users can manage their memory more easily on their own.

Not knowing what is being remembered has been a common concern about memory features. Making the contents of memory visible and letting users tidy it up themselves should make the feature easier to rely on.

Lower Compute Costs Open the Door to Free Users

With Dreaming V3, OpenAI says it cut the computation required for processing to about one-fifth of the previous level. This efficiency makes it possible to extend the feature to free users, who had been hard to reach because of the load, and to expand memory capacity for Plus and Pro users.

The rollout starts with Plus and Pro users in the United States, then widens to more countries over the coming weeks while gradually reaching Free and Go users. Rather than confining an advanced feature to higher tiers, the approach streamlines the foundation first and then delivers it to a broad base of users.

Summary

This update moves ChatGPT's memory from a feature that merely remembers to one that keeps things in order over time. By overhauling the Dreaming memory technology to V3, OpenAI improved the accuracy of carrying memory over, and its internal evaluation showed all three metrics rising together. A memory summary page makes stored information visible, and lower compute costs widen the audience all the way to free users. As AI increasingly remembers personal context over the long term, this looks like an announcement that pushed accuracy, transparency, and breadth of availability forward at the same time.