On June 4, 2026, OpenAI rolled out Dreaming V3, a fundamental overhaul of how ChatGPT remembers and uses information from past conversations. The new memory architecture is rolling out first to Plus and Pro subscribers, with free users expected to gain access within weeks, thanks to a dramatic reduction in computational cost — roughly one-fifth of what the previous system required. This upgrade marks the most significant step yet in OpenAI’s effort to solve a problem that has persisted since the original memory feature launched in 2024.
The Two-Year Memory Problem That Dreaming V3 Finally Addresses
ChatGPT’s memory capabilities have gone through three distinct phases since the company first introduced the feature in April 2024. The original system, called Saved Memory, functioned like a sticky note: users had to explicitly tell ChatGPT what to remember by saying things like “remember this.” It only stored information that users consciously flagged, and anything not written down was lost in the next conversation. Worse, saved memories grew stale over time — a detail about an upcoming trip would remain frozen in the system even after the trip was over, causing ChatGPT to behave as if the user were still traveling.
In April 2025, OpenAI introduced the initial version of Dreaming, which added a background process that could reference past conversation history to supplement the saved memory. Personalization improved noticeably, but as OpenAI itself acknowledged at the time, it was still not sufficiently robust as a standalone memory system. Dreaming V3 changes that completely. It no longer plays a supporting role to saved memories — it becomes the core memory system itself.
OpenAI has published benchmark data that tracks the upgrade across three key evaluation axes. Factual Memory — whether information a user provided in one conversation is available in a subsequent session — improved from 41.5 percent in 2024 to 67.9 percent under Dreaming V0, and now reaches 82.8 percent with Dreaming V3. Preference Reflection — the system’s ability to adjust its recommendations based on stated preferences, such as suggesting vegetarian options after being told a user is vegan — moved from 31.4 percent to 55.3 percent to 71.3 percent. The most dramatic gain is in Information Freshness — whether the system understands that a past event is no longer current. That score jumped from just 9.4 percent in 2024 to 52.2 percent in 2025, and now stands at 75.1 percent.
The freshness improvement is particularly significant for regular users. In 2024, the system could correctly handle temporal updates roughly one time out of ten. Today, it manages the task correctly about three times out of four. Anyone who has used ChatGPT regularly has likely experienced the frustration of telling the assistant about a planned trip to Singapore, only to have it continue referencing that trip weeks after returning. Dreaming V3 automatically updates memories based on the passage of time, rewriting them rather than letting them accumulate as outdated artifacts.
Still, the numbers reveal real limitations. An 82.8 percent factual memory score means roughly one in five conversational contexts still fails to draw on relevant past information. And a 75.1 percent freshness score means one in four temporal updates is missed. The improvement is large, but the system is not yet at the point where users can rely on it completely.
What Dreaming V3 Changes: From Notes to Profiles
The simplest way to describe the shift is that memories are evolving from individual notes into structured profiles. The old Saved Memory system was essentially a bullet-point list — “marine biologist,” “lives in Maine,” “enjoys birdwatching” — with no connection between entries and no awareness of how those facts might change over time. Dreaming V3 works differently. It analyzes conversation history in the background, extracts and synthesizes relevant information, then organizes it into categories such as work, hobbies, travel history, and communication preferences. The result is displayed on a memory summary page that users can review, edit, or instruct the system not to raise certain topics.
OpenAI provided a concrete example in its announcement. When a user asks for advice on purchasing underwater photography equipment, ChatGPT without meaningful memory returns a generic checklist of camera housings, strobes, and lenses. ChatGPT with Dreaming V3, by contrast, recalls that the user previously discussed owning a Sony A1 II, housing it with a Nauticam NA-A1II, and using Backscatter Mini Flash 3 and Inon Z-330 strobes. It then suggests specific compatible products and alternatives based on that existing setup.
That comparison, of course, represents an ideal scenario selected by OpenAI to demonstrate the feature. The real-world accuracy of 82.8 percent means one in five interactions will fail to apply relevant context. And the system’s ability to handle temporal updates, while vastly improved, still leaves a quarter of cases unresolved. Users should expect a meaningful upgrade in personalization, but not perfection.
Computational Cost at One-Fifth: How Dreaming V3 Reaches Free Users
A critical detail in this release is the reduction in computational cost. OpenAI has managed to compress the compute required for Dreaming-based memory to roughly one-fifth of what the previous system demanded. That efficiency gain is what makes the wider rollout possible.
Until now, Dreaming memory was available only to Plus subscribers — who pay $20 per month — and Pro subscribers at $200 per month. OpenAI stated that the quality bar had not been high enough to extend the feature to free users. With Dreaming V3, that barrier has been removed. The company has begun rolling out the update to Plus and Pro subscribers in the United States first. International users and free-tier users are promised access “within weeks,” though the specific rollout schedule and regional order remain unannounced. Paid subscribers can also expect increased memory capacity alongside the upgrade.
The More Memory Improves, the More Questions It Raises
As memory performance increases, the questions about what gets recorded and how it gets used become more pressing. A study published in February 2026 and accepted at the ACM Web Conference examined 2,050 memory entries contributed by 80 ChatGPT users. The findings are striking: 96 percent of those entries were generated unilaterally by the system, not explicitly requested by users. Users had not said “remember this” — ChatGPT had autonomously extracted and stored the information from conversation. Further analysis showed that 28 percent of the memories contained personal data as defined by GDPR, and 52 percent contained psychological insights.
The study examined the older Saved Memory system, not Dreaming V3 directly. But Dreaming’s entire architecture is built on autonomously analyzing conversation history and synthesizing memories — a process that could amplify these tendencies rather than reduce them. When a user casually mentions their health status, family composition, or political views in conversation, those details may be incorporated into their memory profile without explicit awareness or consent.
OpenAI has provided several control mechanisms: the memory summary page for review and editing, an option to disable memory entirely in settings, and the ability to toggle “reference chat history” and “reference saved memory” independently under Settings > Personalization. But the gap between control availability and user awareness remains a real concern. A user may not realize that an offhand remark about their health has been stored and is now shaping ChatGPT’s responses.
Regulatory context adds another layer. The EU AI Act came into force in August 2024 and has been rolling out in phases. Regulations applying to general-purpose AI models began in August 2025, and full application of rules covering high-risk AI systems and transparency obligations is scheduled for August 2026. How the automatic synthesis of memory profiles will be treated under these rules is an open question that regulators will likely need to address soon.
The “Knows You” Competition Is Accelerating
Dreaming V3 signals that the competitive axis in AI assistants is shifting. The battle is no longer solely about the quality of a single response — it is increasingly about the quality of the ongoing relationship between the user and the system. Not having to repeat information that was already shared is a feature many users have wanted for a long time. OpenAI’s benchmarks show steady improvement, and the trajectory is encouraging.
But the same improvement creates a new kind of responsibility. Users now have to think about how much the system knows about them. That is not a problem that a single settings toggle solves. The remaining 17 to 25 percent error rate in memory accuracy, combined with the structural issue that users cannot fully track what the system has recorded about them, means the technology is advancing faster than the governance frameworks around it. Dreaming V3 is a meaningful step forward for personalization. It is also a reminder that memory — in an AI system as much as in human relationships — is never just a technical feature. It is a design decision about what matters, what gets kept, and who gets to decide.
