OpenAI's Dreaming V3 Scores 82.8% on Recall -- and Lost the Feature Power Users Actually Used
OpenAI launched Dreaming V3 on June 4, replacing ChatGPT’s hybrid memory architecture with a single automated background synthesis process. By OpenAI’s own evaluations, the system is substantially more accurate. Factual recall reached 82.8%, preference adherence 71.3%, and time-sensitive accuracy 75.1%. A 5x reduction in compute cost made it viable to extend memory to free-tier users for the first time.
The architectural change that enabled those improvements also removed explicit user control over what the model remembers. That tradeoff is now the subject of a growing backlash from Pro subscribers paying $100 to $200 per month.
What Changed
ChatGPT’s memory architecture had run on two layers since 2025. First: a saved memories notebook — an explicit list of facts users had stated or confirmed, with full visibility and edit access. Second: Dreaming V0, a background process launched in April 2025 that could reference broader chat history as a supplement to the notebook. OpenAI acknowledged at V0’s launch that dreaming alone was “never sufficient as a standalone memory system.”
Dreaming V3 removes the notebook. A single asynchronous background process synthesizes memory from across many conversations simultaneously, captures context that arises naturally in conversation without explicit requests, and updates memories as circumstances change. OpenAI’s stated example: a memory reading “you’re going to Singapore in July” rewrites to “you went to Singapore in July 2026” after the trip ends.
The memory timeline:
| Year | Architecture | Recall Rate (OpenAI) |
|---|---|---|
| 2024 | Saved memories only | ~41.5% |
| 2025 | Saved memories + Dreaming V0 | ~67.9% |
| 2026 | Dreaming V3 (background only) | 82.8% |
These figures are vendor-stated and have not been independently verified.
The Control Problem
For users with straightforward conversational needs, the change is likely invisible or positive. For users who maintained structured, intentional memory states, the notebook disappearing is a direct regression in control.
Under the old system, a user could guarantee specific facts persisted. Under Dreaming V3, the background process determines relevance autonomously. There is no option to pin a memory as permanent. There is no revision history — if the system changes a memory entry, the change is silent. Users can view a synthesized profile summary and mark items as “don’t mention this,” but cannot export, audit, or anchor the underlying state.
A detailed user report surfaced on the OpenAI Developer Community forum on June 8 framing the architecture switch as “cost reduction marketed as a feature upgrade.” The critique: OpenAI reduced compute costs by 5x — making it viable to serve dreaming to free users — at the cost of removing explicit control from paid tiers. The explicit notebook was more expensive to serve at scale and less prone to drift. The automated process is cheaper and more capable on recall metrics, but less trustworthy on continuity.
The Scale vs. Control Tradeoff
OpenAI’s framing emphasises what V3 gained. The system now automatically tracks that you went to Singapore, not just that you were planning to. It handles multi-session, multilayered project context without requiring explicit save actions. For the hundreds of millions of ChatGPT users who never interacted with the saved memories interface, V3 is strictly an improvement.
For the smaller cohort of heavy users who did interact with it, the argument is different. They had a contract with the old system: if you saved it, it persisted. Dreaming V3 has no such guarantee. The background process decides what is “stale” and may revise or deprioritize it without notice.
Users retain a partial disable option — turning off memory prevents new consolidation but does not delete existing stored context or chat history. Full deletion requires deleting individual conversations. The audit interface shows a narrative summary rather than a raw memory state.
OpenAI has not publicly addressed the backlash beyond the launch post. There is no stated plan to reintroduce pinned memories or user-controlled exemptions from auto-revision. The free-tier rollout is continuing as scheduled.