ChatGPT Gets Smarter Memory

OpenAI is deploying a significant upgrade to ChatGPT's memory capabilities, aiming to create a more helpful and context-aware AI assistant. Codenamed 'Dreaming,' the new system is designed to overcome the limitations of previous memory functions, which struggled with staleness and accuracy over time.

The company announced the rollout of a more scalable system for synthesizing memory, tackling issues like outdated information and the challenge of maintaining context across millions of users and multi-year interactions. This update is initially available to Plus and Pro users in the US, with wider availability planned.

Memory's Evolution

Launched in April 2024, the initial 'saved memories' feature allowed users to explicitly instruct ChatGPT to retain information. However, this method was akin to a user taking notes, often missing nuances and becoming stale or incorrect over time. This is a notable evolution from the initial saved memories ChatGPT offered.

A year later, OpenAI introduced the first version of 'dreaming,' enabling ChatGPT to automatically curate memories from chat history in the background. Unlike saved memories, this process synthesized information across multiple conversations, aiming for fresher, more relevant context without explicit user prompts.

While dreaming improved personalization, it wasn't sufficient as a standalone system until now. The latest architecture builds upon dreaming, creating a more capable and compute-efficient memory system. Synthesized memories are now reviewable via a summary page, where users can view, update, or instruct the AI on how to use the information.

Evaluating Memory

OpenAI evaluates memory based on three core objectives: carrying forward useful context, following user preferences and constraints, and staying current over time. The company claims the new dreaming-based system significantly improves recall of relevant facts.

For example, in a complex scenario like purchasing underwater photography gear, a user with memory enabled receives tailored recommendations based on their specific camera setup. Without memory, ChatGPT provides a generic list requiring extensive user research.

Similarly, when planning a trip, ChatGPT with memory can incorporate prior preferences, such as a user's interest in wildlife photography or preference for quiet dinners, leading to more personalized itineraries.