For most of its existence, ChatGPT has been a stateless tool. Each new conversation started fresh, with no recollection of last week's discussion, your current project, or even that you prefer concise answers. OpenAI has been chipping away at this since early 2024, but the release of Dreaming V3 on June 4, 2026 marks the most significant architectural change yet — moving from a manually curated list of facts to a continuously self-updating model of the user.
Here's what changed, how it works technically, and what it means for anyone using ChatGPT for extended work.
The Memory Problem ChatGPT Has Always Had
A language model processes a context window and generates a response. Once the conversation ends, nothing persists unless the application layer explicitly saves it. OpenAI's first attempt at solving this, launched in early 2024, was a simple "saved memories" list: you could tell ChatGPT to "remember" something, and it would store that as a short text snippet to prepend to future conversations.
This worked for simple preferences but broke down quickly for anything dynamic. If you told ChatGPT you were planning a trip to Singapore in July, it would still reference that trip as upcoming in September. The list was static, and keeping it accurate required constant manual maintenance.
In April 2025, OpenAI introduced Dreaming V0, which supplemented the saved list by referencing broader chat context. It was an improvement, but still largely reactive — it didn't proactively update or synthesize information.
What Dreaming V3 Actually Does
Dreaming V3 replaces the previous architecture with a background process that continuously analyzes conversation history to build and maintain a dynamic user profile. The key shift is from explicit storage to auto-synthesis.
Instead of storing isolated facts, the system uses relational embeddings to link information semantically. "User is going to Singapore in July" isn't stored as a standalone note — it's connected to travel patterns, work context, and timeline. When July passes, the system automatically updates the entry to reflect the trip is in the past.
OpenAI describes the system as operating on three core metrics:
- Freshness: Prioritizing recent context over stale information
- Continuity: Connecting threads across time (linking a project mentioned three months ago to a question asked today)
- Relevance: Filtering noise so low-signal information doesn't crowd out high-signal context
The "dreaming" metaphor refers to the background consolidation process — the system reorganizes stored context during idle periods. This allows it to handle long-horizon tasks: tracking a months-long project, noticing when preferences shift, or recognizing that a technical stack you used to ask about has been replaced.
The Performance Numbers
OpenAI published internal evaluation metrics comparing Dreaming V3 to previous iterations:
| Metric | 2024 (Saved Memories) | 2025 (Dreaming V0) | 2026 (Dreaming V3) |
|---|---|---|---|
| Factual Recall | 41.5% | 67.9% | 82.8% |
| Preference Adherence | 55.3% | — | 71.3% |
| Time-Sensitive Accuracy | 52.2% | — | 75.1% |
These are meaningful improvements, particularly in time-sensitive accuracy — whether the system correctly handles facts that change over time. Going from 52.2% to 75.1% suggests the temporal awareness mechanism is doing real work.
The important caveat: these are OpenAI's own internal evaluations, not independently audited. The numbers should be read as directional rather than definitive.
Compute Efficiency and Who Gets Access
Dreaming V3 achieves roughly a 5x reduction in compute cost compared to the previous memory system — what makes the feature viable for free-tier users, where the overhead was previously only justifiable for paying subscribers.
The rollout began June 4, 2026 for ChatGPT Plus and Pro users in the United States, with plans to expand to additional countries and the free tier. Plus and Pro users also received double the memory capacity.
User Control: The Memory Summary Page
The shift to auto-synthesized memory creates a transparency problem that didn't exist with the old list-based system. When you manually told ChatGPT to remember something, you knew exactly what it knew. With Dreaming V3, the system makes inferences — and those inferences might be wrong, outdated, or things you'd rather it not retain.
OpenAI's answer is the Memory Summary Page, a new interface that lets users view, edit, or delete specific inferences, provide explicit instructions about what to retain, and use "temporary chat" mode to opt out of memory storage entirely.
For sensitive information — health data, financial details — the system flags it and requires explicit user approval before storing it. The system can infer sensitive context, but it doesn't retain it without consent.
The Privacy Dimension
The move from raw chat retention to synthesized user models introduces a new privacy consideration. A saved memory ("I have a peanut allergy") is a fact you chose to share. A synthesized inference ("User appears to be managing a chronic health condition based on recurring questions") is something the system derived — and the distinction matters for how users think about what they're sharing.
OpenAI's enterprise implementation addresses this with data isolation, tenant-specific storage, and audit logging for GDPR and HIPAA compliance. For individual users, the Memory Summary Page is the primary control surface. Whether that's sufficient will depend on how transparent the system is about its inferences — something that will become clearer as more users interact with it.
What This Means in Practice
For developers and researchers using ChatGPT for extended work — multi-week coding projects, ongoing research, iterative writing — Dreaming V3 addresses a real friction point. The system should maintain context about your project structure, preferred libraries, and writing style without requiring you to re-establish it at the start of every session.
Tools like MemGPT and Mem0 have been building persistent memory layers for LLMs as standalone products. Dreaming V3 brings similar functionality natively into ChatGPT, which will likely shift the competitive landscape for memory-augmented AI tools.
The open question is accuracy. A memory system that confidently recalls incorrect inferences is worse than no memory at all — it introduces subtle errors that are harder to catch than obvious hallucinations. The 82.8% factual recall figure means roughly one in six facts is wrong or missing. For professional workflows where precision matters, users will need to actively audit the Memory Summary Page rather than trusting the system blindly.
Conclusion
Dreaming V3 is a substantive architectural change, not a marketing update. The shift from explicit list storage to auto-synthesized relational memory, combined with temporal awareness and a 5x compute efficiency gain, represents real engineering work. The performance metrics show meaningful progress, and the Memory Summary Page is a reasonable attempt at giving users visibility into what the system knows.
Whether the inferences are accurate and trustworthy enough for serious work will only be answered by extended real-world use. But the direction is clear: OpenAI is building toward a ChatGPT that functions as a persistent working assistant rather than a stateless question-answering tool.
Primary source: OpenAI — ChatGPT Memory: Dreaming
Supporting sources: DigitalApplied — Dreaming V3 Guide · WindowsForum — Dreaming V3 Architecture · The Cryptonomist — OpenAI Dreaming V3
Top comments (0)