OpenAI gives ChatGPT the ability to remember facts from your chats

When you next converse with ChatGPT and mention your aversion to avocados or preference for bullet point-formatted meeting notes, the chatbot may be able to permanently remember such details. OpenAI recently announced that it is testing an option for users to request the popular chatbot to retain specific information from one conversation to the next. ChatGPT can also autonomously determine which pieces of a user’s discussions should be stored.

Initially, OpenAI will make these features available to hundreds of thousands of free and paid ChatGPT users. The startup plans to gather feedback before expanding the rollout to a wider audience, according to the company’s statement to Bloomberg News. This new memory feature highlights OpenAI’s ongoing efforts to enhance user experience and customization for its 100 million weekly users, all while competing against a growing list of rival chatbots.

The ability to compile a detailed history for each user raises potential privacy concerns, but it also serves as a means for internet companies to personalize each person’s experience and foster customer loyalty. Previously, OpenAI allowed users to provide custom instructions to ChatGPT to avoid repeating their preferences in each chat. However, this new feature takes it a step further.

Joanne Jang, OpenAI’s product lead focusing on AI model response to users, demonstrated the feature’s potential by sharing three different facts with ChatGPT – one in each of three separate conversations. In a fourth chat, Jang asked ChatGPT to create a birthday card for her daughter. The chatbot used OpenAI’s DALL-E 3 image-generator to create a pink card adorned with jellyfish that read: “Happy 5th Birthday Lina.”

OpenAI will notify users if they gain access to the feature. Those with access will have control over the data, being able to delete individual details, erase all of it at once, or choose to turn off the memory option entirely. Additionally, users can ask ChatGPT to provide a summary of any information it has stored.

ChatGPT already avoids requesting sensitive personal information such as passwords or passport numbers. The new feature is designed to refuse to save such data as well, according to Jang. Information saved in ChatGPT’s memory will be treated as standard conversational data and can be used to train the company’s AI models unless users have chosen to exclude their chats from training.

Liam Fedus, a research scientist working on ChatGPT and GPT-4, explained that the memory feature can store a few thousand tokens of written information. OpenAI’s software segments words into smaller components known as tokens to process text. For context, 2,048 tokens is roughly equivalent to 1,500 words, about three times the length of this article, Fedus stated to Bloomberg.