OpenAI Upgrades ChatGPT Memory With Better Recall and Preferences
In a blog post, OpenAI detailed a new memory architecture called Dreaming. The technology works in the background to organise and refresh information gathered from earlier conversations, helping ChatGPT retain relevant context without requiring users to repeatedly provide the same details. The feature is currently rolling out to Plus and Pro subscribers in the US, while support for additional regions and Free and Go users is expected in the coming weeks.
The latest release builds on earlier memory features introduced over the past two years. OpenAI launched Saved Memories in 2024, giving users the ability to ask ChatGPT to remember specific information. A year later, the company expanded the feature by enabling the chatbot to reference past conversations through an earlier version of Dreaming.
OpenAI said the updated system focuses on three areas, including retaining useful context from previous conversations, applying user preferences more consistently, and ensuring stored information remains relevant as circumstances change.
Alongside the announcement, the company published internal benchmark results that suggest improvements across all three categories. Factual recall scores increased from 67.9 percent to 82.8 percent, while preference adherence rose from 55.3 percent to 71.3 percent. Performance on time-sensitive memory tasks also improved, increasing from 52.2 percent to 75.1 percent.
Users will also gain access to a new memory summary page that shows what ChatGPT has learned over time. The page allows users to review stored information, make corrections, remove details, or indicate topics they want the chatbot to remember in future conversations.
The company also reported a substantial reduction in the computing resources required to operate Dreaming. According to OpenAI, these efficiency improvements will help bring the feature to more users while increasing memory capacity for existing Plus and Pro subscribers.
