ChatGPT Dreaming Memory Update

OpenAI has introduced a significantly more capable and compute-efficient memory architecture for ChatGPT based on a process called "dreaming." This system allows ChatGPT to automatically curate and synthesize memory in the background by referencing chat history, reducing the staleness of saved information and improving the model's ability to maintain continuity across multi-year time horizons.

The Evolution of ChatGPT Memory

ChatGPT's memory system has transitioned from explicit, user-triggered saves to an automated background synthesis process to solve challenges regarding scalability and correctness.

• Saved Memories (April 2024): The initial system relied on strong cues (e.g., "remember I am traveling to Singapore") and only wrote memories during active conversations. This approach often resulted in stale information and required explicit instructions to function.
• Dreaming V0 (April 2025): OpenAI introduced the first version of "dreaming," allowing the model to reference chat context outside of the saved memories list and automatically curate memories in the background.
• Dreaming V3 (June 2026): The current architecture is a more scalable and compute-efficient version of dreaming. It serves as a standalone memory system that synthesizes the freshest and most relevant context without relying on explicit user requests.

Users can manage these synthesized memories via a memory summary page, where they can review highlights of what ChatGPT knows, update personal information, and provide instructions on specific topics.

Key Memory Objectives and Capabilities

OpenAI evaluates the effectiveness of the dreaming-based memory system across three primary objectives: carrying forward context, following preferences, and staying current over time.

Carrying Forward Useful Context

The system enables ChatGPT to build on prior context for complex, long-running projects, eliminating the need for users to re-introduce themselves or their technical setups in new chats. For example, if a user has previously discussed a specific camera and housing setup, the model can provide tailored product recommendations based on that specific hardware rather than providing generic checklists.

Following Preferences and Constraints

Dreaming improves the model's ability to apply personal constraints and implicit preferences across different conversations. These preferences include:

• Explicit Instructions: Specific directives on how the model should respond (e.g., "don't bring up [Topic X] again").
• Personal Constraints: Fixed attributes (e.g., "I am vegetarian").
• Implicit Preferences: Contextual data that shapes relevance (e.g., knowing a user lives near San Francisco to tailor local recommendations).

Staying Current Over Time

Unlike traditional memory systems that remain static, dreaming allows memories to be automatically updated as time passes. This prevents "stale memory" where a model might believe a user is still on a trip that ended weeks ago. The system can automatically revise a memory from "The user is planning a trip to Singapore in July" to "The user went to Singapore in July 2026," ensuring that subsequent recommendations are based on the user's current location and time zone.

Scalability and Availability

OpenAI has reduced the compute required to serve the dreaming process by approximately 5x, enabling a wider rollout of the feature.

• Plus and Pro Users: The update is available to users in the US as of June 4, 2026, with increased memory capacity.
• Free and Go Users: The system will roll out to additional countries and to Free and Go users in the coming weeks.