OpenAI Upgrades ChatGPT Memory With New 'Dreaming' Architecture
OpenAI has begun rolling out a significantly upgraded version of its ChatGPT Dreaming memory system, called Dreaming V3, which processes user context automatically in the background. Plus and Pro subscribers in the United States get access starting today, June 4, 2026, with a broader rollout to additional countries and Free and Go users planned over the coming weeks.
How Dreaming Works Differently From ChatGPT's Earlier Memory
ChatGPT's original memory feature, launched in April 2024, required users to explicitly tell the model what to retain. Anything not flagged was lost, and saved memories tended to go stale with nothing to keep them current over time.
A first version of dreaming arrived in April 2025, adding a background process that pulled context from chat history without waiting for a direct command. That version ran alongside saved memories rather than replacing them. Dreaming V3 now works as a standalone system, built to handle hundreds of millions of users across conversations spanning multiple years.
Memories built by dreaming are visible through a new memory summary page, where users can review retained context, make corrections, add information, and direct ChatGPT on what topics to surface or leave out.
Three Memory Problems the New System Addresses
According to OpenAI, the new architecture tackles three problems that older memory versions consistently ran into.
One is context continuity. A user who previously discussed their photography gear can ask for compatible accessories in a new chat and get tailored recommendations without re-explaining their setup from scratch.
Another is preference tracking. Dietary restrictions, communication preferences, and personal constraints all carry forward without needing to be repeated in every new conversation.
Staleness proved harder to fix. Older systems could lock in outdated context. A trip described as "upcoming" might stay that way weeks after the user had already returned home. Dreaming V3 updates memories automatically as circumstances change, correcting time-sensitive context without any prompt from the user.
What the Compute Breakthrough Means for Free Users
Compute cost is what kept dreaming off Free tier until now. Engineering improvements behind V3 cut the compute required to approximately one-fifth of its previous level, according to OpenAI. That's what makes the Free tier rollout possible, alongside an increase in memory capacity for Plus and Pro subscribers.
Users who prefer not to use the feature can turn it off through memory settings, same as before.
The more significant shift is architectural. Most AI assistants have concentrated on expanding what a model can hold in a single session. OpenAI is betting instead on ChatGPT personalization that compounds across months of use. A longer context window helps within one conversation; a memory system that already knows your preferences from six months ago helps across every conversation.
That 5x compute efficiency gain is what makes that bet viable at scale, not just for paying subscribers but across OpenAI's entire user base. Whether people actually want an AI updating its understanding of them in the background, without a direct ask, is the harder question Dreaming V3 still has to answer. For full details on memory controls and how to manage what ChatGPT retains, OpenAI has published a full breakdown of the Dreaming memory update on its website.