OpenAI has begun rolling out a new memory system for ChatGPT, starting with Plus and Pro users in the US.
The revised system is designed to improve how ChatGPT retains and updates information about users across conversations. It aims to address stale information, improve accuracy and scale memory use across a large user base over longer periods.
Memory in ChatGPT stores details such as user preferences, ongoing projects and personal constraints so later conversations can begin with shared context. OpenAI said this has become a significant part of how people use the service over the past two years.
Memory shift
The latest system builds on a background process OpenAI calls dreaming. It lets ChatGPT draw on previous chats to assemble and refresh a user's memory state, rather than relying only on items users explicitly ask it to save.
Earlier versions were more limited. OpenAI introduced saved memories in 2024, allowing users to instruct ChatGPT to retain specific facts for future conversations, but the model depended on direct prompts to store information and could still miss relevant context.
A year later, OpenAI expanded memory by allowing ChatGPT to reference chat context beyond the saved memories list. That marked the first use of dreaming, which curated memories in the background from chat history.
The new architecture builds on that system and is intended to provide a more efficient, scalable foundation. Users can review created memories through a summary page that shows highlights of what ChatGPT knows about them, lets them update details and add instructions about topics the assistant should raise and when.
How it works
OpenAI frames good memory around three tests: carrying forward useful context, following preferences and constraints, and staying current as circumstances change over time. It said it compared the latest version with earlier memory systems introduced in 2024 and 2025.
In practice, carrying forward context means ChatGPT should remember factual information from earlier discussions when it is relevant to a new request. OpenAI gave the example of a user shopping for camera gear, where the model could draw on prior conversations about the person's existing equipment to suggest compatible products.
Preference handling is another part of the update. Memory should help ChatGPT reflect a user's stated or implied preferences, such as dietary choices, response-style instructions or location-based relevance.
OpenAI cited a vegetarian user asking for meal suggestions. Another example is a travel-planning conversation in which the model recalls that a user prefers quiet dinners or wants a hotel with strong air conditioning.
The third issue is time. Older memory systems could become outdated when temporary facts remained active after they were no longer relevant, such as a trip to Singapore that had already ended.
Under the new system, memories can be revised as time passes. A future conversation should therefore reflect that a trip has finished rather than continue to treat it as an upcoming event.
Scale and access
OpenAI said the latest changes also reduce the computing resources needed to run dreaming. Recent improvements cut the compute required to serve the feature to Free users by about five times, according to the company.
That reduction is intended to support a wider rollout beyond paid users. Access will expand to additional countries and to Free and Go users over the coming weeks, while Plus and Pro users will also receive increased memory capacity.
Dreaming now forms a shared memory foundation across OpenAI's user base, it said. The memory summary page is meant to give users more visibility into what the system retains, alongside options to amend or direct that information.
The update comes as AI companies place greater emphasis on persistent personalisation in chatbots, with memory seen as a way to make assistants more useful across repeated interactions. The challenge has been making those systems accurate and current enough to avoid introducing irrelevant or outdated information into new conversations.