ChatGPT's Dreaming V3: How OpenAI Rebuilt Its Memory System

The original ChatGPT memory worked like a sticky note system. You'd mention a preference, the model would write it down in a list, and it would check that list at the start of future conversations. Tell it you're vegetarian and it remembered. Tell it you're applying for jobs and it remembered that too, even six months later when you'd already started a new role.

The problem is obvious in retrospect: sticky notes don't expire. The system stored states, not situations. It had no way to know whether "looking for a new apartment" was still true or whether you'd moved in March. And because the retrieval step searched those stored notes to populate context, stale memories produced stale conversations.

OpenAI's own internal evaluation put a number on this: factual recall accuracy was 41.5% in 2024. More than half the time, the memory system either surfaced the wrong information or retrieved context that didn't help.

Dreaming V3, which began rolling out to ChatGPT Plus and Pro subscribers in the United States on June 4, replaces the sticky note model with something fundamentally different.

What Dreaming V3 Actually Does

The name borrows loosely from neuroscience theories about memory consolidation during sleep. The idea is that the brain doesn't just record experiences verbatim, it integrates them, updates existing knowledge, and prunes what's no longer relevant. Dreaming V3 applies an analogous process to ChatGPT's memory state.

It works as a background synthesis pipeline. After conversations end, a process runs over your interaction history and existing memories. It identifies patterns, resolves contradictions, and updates context that's become outdated. The distinction from the prior system is that this is synthesis, not storage. The model isn't just logging what you said, it's integrating it with what it already knows about you.

OpenAI built the architecture around three axes. Freshness means memories update as circumstances change: "You're going to Singapore in July" becomes "You went to Singapore in July" once that date passes. Continuity means conversational threads pick up coherently across sessions, even when months separate them. Relevance means the retrieval step is smarter about which memories actually matter for a given conversation rather than pulling everything that keyword-matches.

The result in OpenAI's internal evals: factual recall rose from 41.5% to 82.8%. That's a significant jump, even accounting for the self-reported nature of the metric.

The Engineering Trade-Off Worth Noting

Dreaming V3 delivers this improvement while using roughly one-fifth the compute of the prior memory architecture. That matters for OpenAI's unit economics and for the feasibility of rolling this out to tens of millions of users.

The efficiency gain comes from moving work offline. Under the old system, memory search happened during active inference, when compute costs are highest and latency is most visible. Dreaming V3 preprocesses and organizes memories between sessions. The synthesis pass is compute-intensive, but it runs in the background and doesn't add latency to the conversation itself.

This is a pattern showing up across AI infrastructure broadly: expensive, latency-sensitive operations get moved out of the real-time path as the underlying models get more capable. For context on how ChatGPT has changed at the model level recently, see OpenAI's June retirement of GPT-4.5.

The trade-off is freshness lag. Changes from a conversation won't be fully synthesized until after the session ends. For almost all practical use cases, that's fine. The exception would be if you're trying to immediately verify that a preference you just stated is being retained, which the old system handled more literally.

What You Can Still Control

Dreaming V3 doesn't remove user control, it changes how control works. You can still view what ChatGPT has stored about you, and you can edit or delete any memory. The model no longer maintains a simple line-item list that you can audit exhaustively, though. The memory state is more of a synthesized representation than a raw log.

That's a real privacy consideration. Under the prior system, the model's knowledge of you was explicit and enumerable: you could read every entry. Under Dreaming V3, the knowledge is more diffuse. The synthesis process that makes memory smarter also makes it less transparent.

For users who want no persistence at all, temporary chats continue to work exactly as before. Nothing from those sessions gets stored or referenced in future conversations.

Rollout and Availability

The initial rollout covers ChatGPT Plus and Pro subscribers in the United States. No firm timeline has been given for Free tier or international expansion, though OpenAI's pattern is to expand Plus features within a few weeks of initial launch.

ChatGPT Team and Enterprise accounts are not part of the initial release. That's likely deliberate: enterprise deployments have compliance requirements around data retention and auditability, and a system that synthesizes memory rather than storing discrete facts is harder to audit. It's a reasonable call to prove the architecture on consumer accounts first.

If you're on Plus or Pro and based in the US, the new system will build context about you over the next several sessions as the synthesis pipeline runs. It's worth giving it a few conversations before evaluating the difference. The improvements in recall are most noticeable for users who've had consistent ChatGPT usage across multiple months.

For a broader look at how ChatGPT compares to other major AI assistants today, our AI chatbot app guide and ChatGPT's full profile on Chatbot Gallery cover current capabilities across platforms.

What This Signals

Dreaming V3 is the clearest public signal yet that OpenAI is treating long-term memory as a product problem, not just a model capability. The original saved-memories approach was designed to be tractable and auditable. It worked, but it didn't scale well to the way people actually use ChatGPT across months and different life contexts.

The shift to synthesis also raises the competitive bar. If ChatGPT can now maintain coherent long-term context automatically, that pushes every other major AI assistant, including Gemini, Claude, and Copilot, to match it or fall behind on one of the dimensions users increasingly care about: whether the AI actually feels like it knows you.

Keeping up with which AI tools actually matter for your work is a full-time job. The About.chat newsletter does it for you — one email a week, no filler. Subscribe free.