OpenAI Launches Dreaming V3: ChatGPT Memory Now Synthesizes Itself
OpenAI's Dreaming V3 replaces manual memory saves with a background synthesis engine that consolidates context from past conversations automatically. Factual recall jumped from 41.5% to 82.8% across three generations.
ChatGPT’s memory just got a complete overhaul. OpenAI rolled out Dreaming V3 on June 4, 2026 — a background synthesis engine that automatically consolidates everything you’ve told ChatGPT into a persistent, evolving memory layer, with no manual saves required.
The difference from the previous system is architectural. Before Dreaming, ChatGPT saved memories only when you explicitly asked it to — a brittle model that missed most of what mattered. Dreaming runs asynchronously after conversations end, scanning multiple past threads simultaneously, distilling preferences, facts, and context changes into a continuously updated memory state. That state gets injected into the system prompt at the start of every new conversation.
The numbers back up the pitch. OpenAI reports factual recall climbing from 41.5% on v0 to 82.8% on Dreaming V3 — a figure benchmarked on internal evals measuring how accurately ChatGPT retrieves things you’ve previously told it. Preference adherence sits at 71.3%, and time-sensitive accuracy — the hardest category, where the system must track things like “trip to Singapore booked for July” and revise that to “you went to Singapore in July” once the date passes — reaches 75.1%.
There’s a practical reason the Free tier is finally in scope: OpenAI says Dreaming required roughly 5x less compute than its predecessor to serve at scale, making the economics of a Free rollout viable for the first time.
A few things to understand before opting in. Memories persist in a separate data layer — not conversation logs — and survive conversation deletion unless you actively remove them through the new Memory Summary Page. That page lets you view synthesized entries, correct them, or dismiss them. OpenAI’s own documentation notes the page “does not necessarily include everything ChatGPT may remember.” There’s an audit gap: you can see some of what it knows; not necessarily all of it.
The controls are layered. A “Saved Memories” toggle disables synthesis and queues deletion within 30 days. Temporary Chat prevents memory use and update for any session where you want a clean slate. A separate model training setting governs whether your conversations feed training data — distinct from the memory setting.
The 82.8% factual recall figure comes from OpenAI’s own evaluations, not independent benchmarks. Worth noting, though the trajectory from sub-50% to above 80% across three versions represents a genuine jump in utility. That kind of improvement doesn’t happen through tuning — it requires a different architecture, and that’s what Dreaming is.
The compute efficiency gain is the more underappreciated number. A 5x reduction in serving cost means Dreaming can eventually run for 700 million users, not just paying subscribers. If OpenAI’s internal benchmarks hold up under real-world use, this is the most meaningful personalization upgrade in ChatGPT’s history.
Dreaming V3 is live now for Plus and Pro subscribers in the US. Free-tier users and international markets follow in the coming weeks.