From notes to dreams: How ChatGPT is learning to remember more like a person

OpenAI is rolling out a significantly improved memory system for ChatGPT, designed to give the AI a more accurate and up-to-date understanding of individual users over time. OpenAI writes in their post that the update, built on a technology called “dreaming,” is now available to Plus and Pro users in the United States, with a broader rollout to Free and Go users planned for the coming weeks.

Memory in ChatGPT has existed since April 2024, when OpenAI introduced a feature that let users ask the AI to save specific pieces of information. That system had clear limitations: it only stored what users explicitly told it to remember, and saved memories frequently became outdated.

What dreaming does differently

Dreaming is a background process that reviews a user’s chat history and continuously updates what ChatGPT knows about them. Unlike the earlier system, it does not wait for an explicit instruction. Instead, it synthesizes context from many conversations automatically.

OpenAI identifies three goals for a well-functioning memory system:

• Carrying forward useful facts, such as a user’s camera equipment or dietary preferences
• Applying personal preferences and constraints consistently across conversations
• Staying current as time passes, for example updating “planning a trip to Singapore” to “went to Singapore” once the trip is over

Users can review a summary of what ChatGPT has stored about them, make corrections, and instruct the AI on which topics to raise or avoid.

OpenAI says recent technical improvements reduced the computing resources required to run dreaming by approximately five times, making it practical to extend the feature to Free users. The company describes this update as its most capable memory system to date and says further improvements are planned.