OpenAI launched a new background memory architecture called Dreaming that automatically synthesizes and updates what ChatGPT knows about you across every conversation. For marketers, operators, and anyone who has spent time re-explaining their business context to an AI tool, the overhead just got a lot smaller.
Every professional who uses ChatGPT regularly knows the invisible tax: the two minutes at the start of each session explaining what you do, who you are, what tone you need, and what context the model is missing before it becomes useful. Multiply that by a team of twenty running AI-assisted workflows five days a week, and you are looking at dozens of hours a month spent onboarding a tool that should already know you. On June 4, OpenAI began rolling out Dreaming, a background memory architecture designed to eliminate exactly that friction.
Dreaming is not a new feature so much as a new engine under an existing one. ChatGPT has had some form of memory since April 2024, first as explicit saved notes you had to ask the model to store, then as a supplementary synthesis layer called Dreaming V0 launched in 2025. What shipped this week is Dreaming V3, a substantially more capable and compute-efficient version that replaces the saved-memories list entirely as the foundation of what ChatGPT knows about you.
What Changed, Specifically
The old system required a strong cue. You had to say something like "remember I'm working on a Q3 launch" for ChatGPT to store it. That approach worked for explicit instructions but missed the natural accumulation of context that happens across dozens of conversations over months. It also went stale. If you told ChatGPT you were planning a trip and then took it, the model would still treat you as a person about to travel.
Dreaming V3 runs in the background and synthesizes memory automatically from your conversation history, without requiring any explicit instruction. It resolves staleness by treating memory as time-aware: a note that read "planning a product launch for October" gets updated once October has passed. The system understands that context ages.
OpenAI reports that factual recall improved from 41.5% on their internal evaluation in 2024 to 82.8% in 2026 with Dreaming V3 active. Preference-following and time-sensitive scores sit in the low-to-mid 70s. Those are not perfect numbers, but they represent a substantial change in what the tool actually delivers across a workday.
The new system also includes a memory summary page where users can see what ChatGPT has synthesized, correct errors, add context, or specify what topics the model should and should not surface. That transparency layer matters: memory that operates without any audit trail creates a different kind of trust problem.
The Business Translation
The use case that will matter most to marketing and ops teams is not the dramatic one. It is the low-grade version of the problem that shows up in every session: the model does not know your brand tone, your audience, your competitive constraints, your stack, or your company's internal language unless you re-establish it each time.
Teams that have built around ChatGPT for content drafting, competitive research, meeting prep, or copy review have typically handled this by maintaining a pasted context block that gets dropped into every session. Some use system prompts in the API. Others have team members who just rewrite the same briefing paragraph from memory every morning. Dreaming V3 is a system-level replacement for that workaround.
Practically, this means a marketing leader who has been using ChatGPT to draft campaign briefs no longer needs to re-establish that the company sells mid-market B2B software, avoids jargon, targets RevOps buyers, and wants a specific document structure. A founder who has described their product ten different times across ten different conversations should find the model starting to understand the product as a coherent thing rather than a fresh topic each session.
For agencies running AI-assisted client work, the implication is more structural. Agency work generates dense context across clients, briefs, brand guidelines, and live campaigns. Any workflow that currently starts with a context dump gets faster when the model already carries that context from prior sessions.
Who Gets It and When
Dreaming V3 is currently rolling out to Plus and Pro subscribers in the United States. Free and Go tier users, along with international accounts, will follow over the coming weeks. Enterprise and API customers get access via a new memory_context parameter, which allows developers to build applications that leverage long-term user profiles without storing sensitive data on their own servers.
The rollout is gradual, which means not every paid subscriber has access today even if their plan technically qualifies. OpenAI confirmed the timeline but has not given specific dates for each tier.
The Honest Part
Dreaming V3 improves memory, but it does not solve every memory problem. The recall rate of 82.8% means the model still misses context roughly one in six times. For high-stakes use cases, where a brand voice violation or an incorrect product assumption creates real cost, a background memory system is not a substitute for explicit review. You still need to check what the model thinks it knows.
There is also a precision question. Synthesized memory works well when your context is consistent across conversations. If you use ChatGPT for widely different tasks, the synthesized memory may blend contexts in ways that produce less useful results rather than more. A founder who uses ChatGPT for investor memos, personal travel planning, and technical documentation in the same account is giving the system a harder problem than one who uses it almost entirely for a single domain.
Finally, the transparency tooling is new. The memory summary page that lets users review and correct synthesized memory is a thoughtful addition, but building the habit of auditing it is something teams will need to actually do. Memory that is wrong and invisible is worse than no memory at all.
The Deeper Shift
What Dreaming V3 represents, more than any individual feature, is OpenAI acknowledging that ChatGPT's job is not to be useful in a session. It is to be useful over time, across sessions, for a person who has a life and a context that does not reset every chat window.
That framing has implications beyond memory. It is the same framing that drives ZoomMate's bet on meetings as the anchor for enterprise context, and Microsoft Scout's bet on always-on agents that carry organizational awareness. The AI tool that knows you well enough to skip the briefing is not just faster. It is, for most knowledge workers, categorically more useful.
OpenAI has 500 million weekly users. The gap between what those users have to explain each session and what the model should already know is enormous. Dreaming V3 starts closing it.
The re-explanation tax was always a sign that something in the system design was wrong. Not a user behavior problem, not a prompt engineering problem. A design problem. The fix arrived this week. It is not complete, but it is pointed in the right direction.