OpenAI Rolls Out Smarter ChatGPT Memory With Dreaming

The Story

ChatGPT memory is moving from a mostly manual saved-memory system to a more automatic background process.

Dreaming was developed to address three problems that become harder at large scale: stale information, incorrect context, and the cost of supporting memory across hundreds of millions of users over multi-year time horizons.

Memory first launched in April 2024 as saved memories. That system let users ask ChatGPT to remember specific details and carry them into future chats. In April 2025, OpenAI added the first version of Dreaming, allowing ChatGPT to reference chat context outside the saved memories list. The 2026 update introduces a more capable and compute-efficient ChatGPT memory architecture built on Dreaming.

The Facts

• OpenAI announced the Dreaming memory update on June 4, 2026.
• The rollout starts with ChatGPT Plus and Pro users in the US.
• The feature will expand to additional countries and to Free and Go users over the coming weeks.
• Memory first launched in April 2024 as saved memories.
• The first version of Dreaming was introduced in April 2025.
• The 2026 version is described as Dreaming V3.
• Dreaming uses a background process to synthesize memory from many conversations.
• The system is designed to carry forward useful context, follow user preferences and constraints, and stay current as time passes.
• Users get a Memory Summary page where they can review highlights, add or update information, and guide when certain topics should be brought up.
• Recent improvements reduced the compute required to serve Dreaming to Free users by about 5x.
• Plus and Pro users are expected to receive increased memory capacity.
• Turning off Reference saved memories also turns off Reference chat history.
• Turning off Reference chat history deletes information ChatGPT remembered from past chats within 30 days.
• Temporary Chats do not appear in chat history, do not use memory, and are not used to improve OpenAI’s models.

What Is New

The main change is that ChatGPT memory can now build and update a broader memory state without depending only on explicit saved memories.

Saved memories relied on strong cues, such as a user saying “remember this.” Dreaming can include context that appears naturally in conversation, making it more useful for long-running projects, personal preferences, writing style, recurring tasks, travel plans, professional context, and user constraints.

A simple example: if a user planned a trip to Singapore in July 2026, a stale memory system might keep treating that trip as upcoming after July has passed. A time-aware memory system can update the context so the trip becomes part of the user’s past history instead of a current plan.

How It Works

Dreaming does not mean ChatGPT is “thinking” while a user is away in a human sense. It refers to a background synthesis process that reviews relevant chat history and updates ChatGPT memory so future conversations can start with better context.

Good memory has three main goals:

• carrying forward useful context
• following preferences and constraints
• staying current over time

The goal is not to remember everything. The goal is to keep useful, relevant, and current information available when it improves the answer.

That may include:

• user preferences
• recurring projects
• professional context
• writing or communication style
• location-sensitive or time-sensitive details
• constraints the user often repeats
• topics the user wants avoided or handled carefully

User Controls

The Memory Summary page is the main user-facing control.

Users can review highlights of what ChatGPT knows about them, add or update information, correct details, and provide instructions on what topics ChatGPT should bring up and when.

However, the summary is not a complete memory database. Some details may not appear in the summary, including information considered less relevant or not suitable for that view.

Users can also manage ChatGPT memory through two related settings: Reference saved memories and Reference chat history. Turning off Reference saved memories also turns off Reference chat history. If Reference chat history is turned off, information remembered from past chats is deleted from OpenAI’s systems within 30 days.

Users who do not want information used for future personalization can use Temporary Chats. Temporary Chats do not appear in chat history, do not use or update memory, and are not used to improve OpenAI’s models.

Privacy and Data Considerations

The update makes ChatGPT more useful, but it also makes memory management more important.

Because ChatGPT can personalize responses from past chats, users should periodically review their Memory Summary, remove outdated details, and use Temporary Chat for sensitive or one-off conversations.

Saved memories are stored separately from chat history. Deleting a chat does not automatically remove saved memory from that conversation. To fully remove a saved memory, users need to delete both the memory and the chat where the information was originally shared. OpenAI may retain logs of deleted saved memories for up to 30 days for safety and debugging.

Memory controls and model-training controls are also separate. Users who want to limit both persistent memory and training use need to check both settings.

Best Use Cases

The upgrade is most useful for people who use ChatGPT repeatedly across the same work or personal areas.

Strong use cases include:

• long-running work projects
• writing style consistency
• study plans and learning preferences
• travel planning
• shopping recommendations based on known equipment or preferences
• content calendars
• coding projects
• business strategy work
• personal productivity workflows
• recurring research tasks

For casual one-off questions, the change may be less visible. For regular users, it could reduce the need to re-explain context at the start of every conversation.

Why This Matters

ChatGPT memory is becoming one of the most important features in consumer AI. A chatbot that remembers useful context can become less like a search box and more like a persistent assistant for work, learning, planning, and creative projects. The trade-off is that users need clearer controls, better visibility, and stronger habits around reviewing what the system remembers.

This article was drafted with the assistance of generative AI. All facts and details were reviewed and confirmed by an editor prior to publication.