OpenAI rolls out new ChatGPT memory system to keep personalization current

AI

Plus and Pro users in the US are getting a more current memory summary first, with Free and Go users set to follow over the coming weeks.

A blue-toned digital image showing a human profile layered with city lights, data panels, and connected interface elements, representing ChatGPT memory, personalization, and context retrieval across past conversations

OpenAI has started rolling out a new ChatGPT memory system for Plus and Pro users in the United States

OpenAI has started rolling out a new ChatGPT memory system for Plus and Pro users in the United States, changing how the AI assistant carries context across long-running conversations, projects, and preferences.

The update introduces a more capable version of OpenAI’s "dreaming" memory architecture, which synthesizes useful context from past chats rather than relying only on individual saved memories. OpenAI says the system is designed to improve freshness, continuity, and relevance as ChatGPT is used across multi-year projects and repeated workflows.

The rollout begins in the US for Plus and Pro users and will expand to additional countries, Free users, and Go users over the coming weeks. The change is relevant for education, work, and research users who use ChatGPT across multiple tasks and expect the product to remember constraints, preferences, and project history without being manually updated every time.

The new architecture also makes memory more practical to serve at scale, after recent improvements reduced the compute required for dreaming by approximately five times.

Users can review and correct memory summaries

The new system gives users a memory summary page, where they can review a high-level view of what ChatGPT knows about them. The summary can include work, hobbies, preferences, travel interests, projects, and other recurring context.

Users can add or update information from the memory summary, correct specific details, or tell ChatGPT not to mention something again. OpenAI says the summary is automatically updated as people continue using ChatGPT, although it may not show every detail used to personalize responses.

Memory controls are available in Settings > Memory. Users can turn memory on or off, use Temporary Chats, ask ChatGPT what it remembers, or ask it to forget something.

OpenAI also says that fully removing a detail may require deleting it from every place it appears, including past chats, archived chats, files, the memory summary, and connected apps. Temporary Chats do not use existing memories or create new memories.

Memory sources show why a response was personalized

OpenAI is also adding memory sources, which show what information was used to personalize a response. Depending on the user’s plan and region, sources may include past chats, saved memories, custom instructions, files, and connected Gmail.

Memory sources are available on the web for ChatGPT Plus, Pro, Free, and Go users, with mobile rollout planned. OpenAI says sources are designed to make personalization easier to understand and manage, but they may not show every factor that shaped a response.

Files and Gmail sources are available to Plus and Pro users in some regions, but not in the European Economic Area, Switzerland, or the UK. Gmail connections are available on Plus and Pro plans where supported, allowing ChatGPT to use relevant inbox information, such as travel plans, project threads, and scheduling context, when the user has connected Gmail.

ChatGPT Enterprise workspace owners can turn memory on or off for all users in Admin Settings. OpenAI says Reference Chat History is not yet available to Enterprise and Edu customers.

OpenAI reports stronger recall and preference handling

OpenAI says the updated memory system improves how ChatGPT carries forward facts, follows user preferences, and stays current as time passes.

In OpenAI’s factual recall evaluation, task success increased from 41.5 percent for 2024 saved memories to 67.9 percent for the 2025 saved memories and Dreaming V0 system. The 2026 Dreaming V3 system reached 82.8 percent.

For preference adherence, OpenAI reports task success of 31.4 percent in 2024, 55.3 percent in 2025, and 71.3 percent in 2026. In a time-sensitive evaluation designed to test whether ChatGPT updates outdated context, such as a trip that has already ended, task success rose from 9.4 percent in 2024 to 52.2 percent in 2025 and 75.1 percent in 2026.

OpenAI first launched saved memories in April 2024. The 2025 version added the first use of dreaming, allowing ChatGPT to reference context beyond the saved memories list. The 2026 update makes dreaming the core architecture for a wider memory rollout.

The new memory system is available now to Plus and Pro users in the US. OpenAI says more countries and Free and Go users will be added over the coming weeks, with further improvements planned as memory becomes a shared foundation across ChatGPT plans.

Previous
Previous

ETIH Innovation Awards Winners: IXL receives Editor’s Choice Award

Next
Next

Microsoft says agentic AI helped make its new quantum chip 1,000 times more reliable