ChatGPT's Dreaming V3 Rebuilds Memory From the Ground Up — and Free Users Finally Get It Too

ChatGPT’s memory has come a long way from its awkward beginnings, but until now it has always felt like a feature in progress rather than a finished one. On 4 June 2026, OpenAI shipped Dreaming V3, a rebuilt memory architecture that replaces the patchwork of saved memories and early dreaming experiments with a single, coherent system. It is more accurate, more time-aware, cheaper to run, and for the first time, available to free-tier users.

A brief history of ChatGPT memory

OpenAI launched memory in April 2024 as a simple saved-memory system. You had to prompt ChatGPT fairly explicitly to store a fact, and over time those stored facts went stale. The model had no real mechanism for revising what it knew as your circumstances changed.

In April 2025, OpenAI introduced “dreaming,” a background process that let ChatGPT review past conversations and synthesise memories without waiting for you to explicitly save anything. It was a genuine improvement, but by OpenAI’s own account it was never strong enough to stand alone as the primary memory system. It supplemented saved memories rather than replacing them.

Dreaming V3 changes that. It is now the primary system, not a supplement.

What has actually changed

The headline numbers from OpenAI’s internal evaluations tell the story clearly.

Factual recall improved from 41.5% (2024 saved memories) to 67.9% (2025 dreaming) to 82.8% with Dreaming V3. Preference adherence followed a similar curve, rising from 31.4% to 55.3% to 71.3%. The most dramatic improvement is in time-sensitive memory, where accuracy climbed from just 9.4% with the original system to 75.1% today.

That last one matters more than it might sound. One of the persistent frustrations with AI memory is that context goes stale. If you told ChatGPT in June that you were travelling to Singapore in July, it would still be referencing that future trip in August. Dreaming V3 is designed to revise its understanding as time passes, updating “you’re going to Singapore in July” to “you went to Singapore in July 2026” once the date has passed, and then adjusting recommendations back to your home context accordingly.

The system also doubles memory storage capacity for Plus and Pro users.

Why free users are getting access now

The original dreaming architecture was compute-intensive. Serving it to hundreds of millions of free accounts was not practical. OpenAI says recent engineering work cut the compute required by approximately 5 times, crossing the threshold where rolling it out to free users became both technically feasible and economically sensible.

Plus and Pro users in the US are seeing the upgrade from today. Free and Go users will follow over the coming weeks as the rollout expands globally.

The memory summary page

Rather than giving you a raw list of stored facts, Dreaming V3 surfaces what it knows through a memory summary page. You can read a plain-language overview of what ChatGPT understands about you, see when it was last updated, and edit it directly by typing into a text box or highlighting specific sections.

One clarification worth noting: selecting “Don’t mention this again” on a highlighted item suppresses references to that information but does not delete it. If you want something genuinely removed, you need to delete every source where it appears: past chats, archived chats, files, the memory summary itself, and any connected apps. That is a more involved process than it might initially appear.

The summary also works alongside memory sources, a feature OpenAI released with GPT-5.5 Instant that shows you exactly which context was used to personalise a given response.

What this means for you

If you use ChatGPT regularly for anything that benefits from continuity, such as ongoing projects, writing in a consistent style, planning that spans weeks or months, or simply not having to re-explain your preferences every session, Dreaming V3 should feel like a meaningful upgrade. The improvements to time-aware memory in particular address a real daily frustration.

For free users, this is the first time the background synthesis system has been available at all. You will not get the doubled storage that Plus and Pro users receive, but you will get the same underlying architecture, which is a more substantive upgrade than the tier difference might suggest.

If you would rather retain the older saved-memories behaviour, you can switch back in settings. Temporary Chats remain an option if you want a session that neither reads from nor writes to memory.

A note on privacy

As this feature reaches a much larger user base, it is worth being clear-eyed about what memory involves. A February 2026 academic paper found GDPR-defined personal data in 28% of memories and psychological insights in 52% of memories in their dataset. The vast majority of those memories were created by the system without explicit user action.

OpenAI’s guidance is straightforward: if you do not want information from a conversation to be used for personalisation, use Temporary Chats or turn memory off entirely. That is good advice to take seriously before sharing anything sensitive, particularly for users who are new to the feature.

The bigger picture

Memory is where the personal assistant ambitions of AI products either succeed or fall apart. Google, Anthropic, Microsoft and Apple are all working on similar problems. OpenAI’s decision to invest significantly in the compute efficiency of this system, to the point where free users can benefit from it, suggests the company sees persistent, accurate memory as a core differentiator rather than a premium add-on.

Dreaming V3 is a significant step forward. The accuracy improvements are real, the time-awareness problem is finally being addressed properly, and broader access means more people can benefit from ChatGPT actually knowing who they are across sessions. The question now is how quickly the remaining accuracy gaps close, and whether users will trust the system enough to let it accumulate the context that makes it genuinely useful.