Dreaming: An Enhanced Memory System for a More Useful ChatGPT

OpenAI is introducing a more powerful and scalable memory synthesis system to optimize ChatGPT's memory freshness, continuity, and relevance. This update was developed to address the problem of information becoming outdated and inaccurate when handling hundreds of millions of users and years of conversation history.

The existing Saved memories (launched April 2024) required users to explicitly ask ChatGPT to remember something. However, this approach had the limitation that information could become inaccurate or lose relevance over time.

To address this, the Dreaming technology that was introduced automatically curates memories in the background by referencing conversation history. The newly unveiled architecture builds on the existing Dreaming to provide even more efficient and powerful memory synthesis capabilities.

The key features of the new system are as follows:

• Automatic memory synthesis: Naturally learns information through conversational context without requiring explicit requests.
• Staying current: Background processes ensure the freshest and most relevant context is always provided.
• User control: Through the Memory summary page, users can review, edit, or delete the information ChatGPT knows.

This feature is currently being rolled out first to Plus and Pro users in the US, and is expected to expand to other countries and to Free and Go users in the future.