OpenAI is transitioning ChatGPT from a system of reactive "saved memories" to a proactive background synthesis architecture called Dreaming V3. This shift addresses the fundamental engineering challenges of memory staleness, correctness, and scalability when managing context for hundreds of millions of users across multi-year time horizons.
Moving beyond explicit instructions
Since its launch in April 2024, ChatGPT's memory relied heavily on "saved memories," a method where the model only wrote information to a list when triggered by strong cues, such as a user explicitly saying, "remember I'm traveling to Singapore in July." This approach functioned like a person taking sporadic notes; if a detail wasn't explicitly flagged, it was effectively lost. Furthermore, these static entries tended to go stale, becoming irrelevant or incorrect as user circumstances changed.
To solve this, OpenAI introduced "dreaming" in April 2025, a method for the model to automatically curate memories in the background by referencing chat history. While the initial version (V0) supplemented saved memories, it wasn't sufficient as a standalone system. The newly released Dreaming V3 architecture is significantly more compute-efficient and capable of synthesizing a cohesive memory state without requiring constant user intervention.
Synthesizing context for long-running projects
Effective memory must satisfy three core objectives: carrying forward useful context, following persistent preferences, and staying current over time. For example, a user discussing a complex photography setup can later ask for compatible gear using vague terms like "my photography setup," and the model can provide tailored recommendations based on previously synthesized context.
This background synthesis allows ChatGPT to capture nuances that occur naturally in conversation, rather than relying on the user to act as a manual data entry clerk. Users can still review, add, or update these synthesized memories through a dedicated summary page, providing a layer of control over the automated process. This architecture enables ChatGPT to maintain continuity for complex, long-running projects, ensuring that the shared context evolves alongside the user's actual needs.
This deployment marks a significant step toward highly personalized AI agents that maintain a coherent understanding of user identity and intent over extended periods.
Source: Dreaming: Better memory for a more helpful ChatGPT
Domain: openai.com
Comments load interactively on the live page.