ChatGPT ‘Dreaming V3’ Memory: Self-Updating AI Recall

ChatGPT memory just changed shape. On June 4, 2026, OpenAI began rolling out Dreaming V3, a background process that synthesizes memory automatically and, for the first time, replaces the saved-memories list as ChatGPT's standalone foundation. You no longer have to tell ChatGPT what to remember — context that surfaces naturally in conversation is captured, and existing memories update themselves as time passes.

That is a genuine step up from the saved-memories model OpenAI first shipped in April 2024, where you had to explicitly say "remember this" and entries went stale the moment the world moved on. It is also a new kind of responsibility. When a system silently rewrites what it knows about you, the question stops being "did I tell it that?" and becomes "did it revise that correctly, and would I even notice if it didn't?"

This guide covers exactly what Dreaming V3 ships, how the new model differs from saved memories, what OpenAI's own evaluation numbers do and don't tell you, how the feature compares with Claude and Gemini memory, and what self-updating recall means for marketers building personalization on top of these platforms. Where a number is vendor-stated rather than independently verified, we say so plainly.

OpenAI published the announcement, "Dreaming: Better memory for a more helpful ChatGPT," on June 4, 2026. The core change is architectural. Earlier memory worked from an explicit list: saved memories (April 2024) only recorded what you told it to during an active conversation, and Dreaming V0(April 2025) supplemented that list by referencing chat context beyond it — but OpenAI was candid that V0 was never "sufficient as a standalone memory system."

Dreaming V3 closes that gap. It is built on top of dreaming and no longer relies on the saved-memories list as primary storage. A background process learns from many conversations and synthesizes ChatGPT's memory state continuously, capturing context that arises naturally rather than waiting for a "remember this" instruction. OpenAI frames it as "our most capable memory system yet" and a "shared memory foundation for all users" — language that signals memory is now a platform primitive, not a power-user convenience.

The cleanest way to understand the shift is by what each system asks of the user. Saved memories asked you to curate — to decide what was worth keeping and to remember to update it. Dreaming asks nothing, and that is precisely what makes it both more useful and harder to supervise.

OpenAI's own framing of the design goal is worth quoting in full, because it names the three problems this release exists to solve:

Staleness was the known design flaw in saved memories: an entry written during one conversation and never revisited would eventually become wrong — a past location read as current, a resolved project treated as live, an outdated preference still steering responses. Dreaming V3's temporal awareness targets exactly this. OpenAI's canonical example is a memory that reads "the user is going to Singapore in July" automatically rewriting itself to "the user went to Singapore in July 2026" once the trip has passed, without any prompt from the user.

That is the conceptual leap. Memory is no longer a snapshot you have to keep accurate; it is a living state the system keeps current on your behalf. The cost of that convenience — who notices when an automatic revision is wrong — is the thread we return to in Section 08.

OpenAI defines "good memory" against three objectives: carry forward useful context, follow preferences and constraints, and stay current over time. It evaluates Dreaming V3 against all three using 2024 (saved memories), 2025 (saved memories plus Dreaming V0), and 2026 (Dreaming V3) as reference points. The direction of travel is the real story — and it is a steep one.

On factual recall, OpenAI reports improvement from roughly 41.5% in 2024 to about 67.9% in 2025 to 82.8% in 2026. Preference adherence reaches the low 70s with Dreaming V3, and time-sensitive accuracy lands in the mid 70s. Before you anchor on those figures, the important caveat: these are OpenAI's own internal evaluations, and the specific percentages come from secondary coverage reading them off the announcement's charts. OpenAI did not publish the methodology, dataset, or reproducibility instructions.

What the trend does tell you is meaningful even with the caveats. Carrying context forward — the most basic job of a memory system — went from failing most of the time to succeeding most of the time across two release cycles. The two dimensions that lagged in older systems, following stated preferences and staying current, are now the ones OpenAI is explicitly optimizing. That is a coherent engineering arc, and it is the arc marketers should plan around even if the exact decimals are soft.

Because memory now updates without your input, OpenAI paired the release with a transparency surface: the Memory Summary Page. It shows a high-level overview of what ChatGPT knows about you — work, hobbies, travel interests, community involvement — and gives you direct controls over each element.

This is a real improvement in legibility, and it is also where the responsibility quietly shifts. The Memory Summary Page is the answer to "how do I see what it knows?" — but it makes auditing a user task. The system corrects itself automatically; confirming those corrections were right is on you. For teams building products and campaigns on top of ChatGPT, that distinction matters, and we unpack it in Section 06.

Dreaming V3 lands in a market where every major assistant now has a memory story — but they are architecturally distinct, and the differences map onto real workflow tradeoffs. ChatGPT bets on implicit, auto-synthesized global memory. Claude leans on explicit, user-controlled, project-scoped memory. Gemini ties memory to a distilled profile and the wider Google ecosystem. The matrix below is our own consolidation across the three platforms' current documentation and reporting.

A few specifics behind the matrix. Anthropic rolled out persistent Chat Memory across all Claude plans — including free — earlier this year, and paired it with the ability to import memories from ChatGPT, Gemini, and Grok, which removes a meaningful switching barrier. On the Gemini side, paid users have reportedthe assistant forgetting instructions after as few as 25 to 30 messages in active conversations — a context-management complaint rather than a documented limitation, and not confirmed by Google, but a useful contrast with dreaming's background synthesis. If you are weighing a move between platforms, our guide on switching from ChatGPT to Claude without losing your memory covers the portability mechanics in detail.

For anyone building on top of these assistants, self-updating memory cuts both ways. The upside is durable, low-friction personalization: a user's stated preferences, context, and constraints persist and stay fresh across sessions without re-onboarding, which makes assistant-mediated experiences feel continuous rather than transactional. The catch is that you no longer fully control the record — the platform infers, synthesizes, and revises it, and a corrected memory can change behavior in ways you didn't trigger.

The strategic read: memory is shifting from a feature users opt into to infrastructure they live inside. That rewards brands that treat assistant personalization as a first-class channel and punishes ones that assume the user's remembered context is static. If you are mapping how AI memory systems are architected before you build on them, our deeper explainer on how AI memory systems are architected is the engineering companion to this piece, and our content engine work increasingly accounts for how assistants remember and resurface a brand across sessions.

A capable, self-updating memory operating across hundreds of millions of users raises the governance stakes, so the defaults matter. Memory is on by default for Free, Plus, Pro, and Team accounts, and off by default for Enterprise and Edu — where account owners can disable it at the organization level at any time. If you want a session with no memory at all, Temporary Chat mode suppresses it entirely: temporary chats don't appear in history, don't use memory, aren't used to train models, and are deleted after 30 days.

Here is the angle most coverage misses. Dreaming V3 solves staleness by making memory self-correct — but that trades one problem for a subtler one. Under saved memories, a wrong entry was wrong because the system never touched it; you could reason about it. Under dreaming, a memory that was accurate three months ago may have been automatically "corrected" in a way you never explicitly saw. The failure mode moves from neglect to silent revision.

That is a new UX responsibility, and the Memory Summary Page is both the mitigation and the proof of it. The system can now keep your memory fresh; it cannot guarantee every inferred correction is right, and it puts the burden of catching bad revisions on the user rather than on the platform. For privacy-conscious users, the surface to audit is no longer just "what did I tell ChatGPT" — it is "what did ChatGPT infer, and then quietly change."

Looking forward, this is where memory architecture and product design converge. As implicit, self-updating memory becomes the default substrate across assistants — and as it reaches free tiers at near-universal scale — the differentiator won't be whether an assistant remembers, but how legibly it lets you supervise what it chose to remember and revise. Expect the next round of competition to be fought on transparency and correction tooling, not raw recall percentages. If you want the taxonomy underneath all of this, our primer on episodic, semantic, and procedural memory in AI agents explains where dreaming fits in the broader picture.