OpenAI gives ChatGPT better memory.
The San Francisco artificial intelligence startup said Tuesday it will launch a new version of its chatbot that will remember what users said so it can use that information in future conversations.
If a user mentions that his daughter, Lena, who is about to turn five, likes the color pink and enjoys jellyfish, for example, ChatGPT can store and retrieve that information as needed. When the same user asks the bot to “Create a birthday card for my daughter,” it might create a card with a pink jellyfish that says “Happy 5th birthday, Lena!”
With this new technology, OpenAI continues to transform ChatGPT into an automated digital assistant that can compete with existing services like Apple's Siri or Amazon's Alexa. Last year, the company allowed users to add personal instructions and preferences, such as details about their jobs or the size of their family, which the chatbot must consider during each conversation. Now, ChatGPT can draw on a broader and more detailed set of information.
“We believe the most useful assistants are those that evolve with you — and keep up with you,” said Joan Zhang, an OpenAI product lead who helps oversee its memory project.
Although ChatGPT can now remember previous conversations, it can still make mistakes, just as humans do. When a user asks ChatGPT to make a birthday card for Lina, the chatbot may generate one with a subtle typo like “Happy 5th birthday! Lina!”
The company first makes the new technology available to a limited number of users. It will be available to people who use the free version of ChatGPT as well as those who subscribe to ChatGPT Plus, a more advanced service that costs $20 per month.
OpenAI is also offering users on Tuesday what it calls ephemeral chats, during which conversations and memories are not stored.
ChatGPT has for some time provided a limited form of memory. When users chatted with the bot, its responses were based on what they had said previously in the same conversion. Now, the bot can rely on information from previous conversations.
(The New York Times sued OpenAI and its partner Microsoft in December for copyright infringement on news content related to AI systems.)
The robot builds this memory by automatically identifying and storing information that may be useful in the future. “We rely on the model to determine what may or may not be relevant,” said Liam Fidos, an OpenAI research scientist, referring to the AI technology that powers ChatGPT.
Users can tell the bot to remember something specific from their conversation, ask what's already stored in its memory, tell the chatbot to forget certain information or turn off memory entirely.
By default, OpenAI records entire ChatGPT conversations and uses them to train future versions of the chatbot. OpenAI said it has removed personally identifiable information from conversations used to train its technology. Users can choose to remove their conversations from the OpenAI training data entirely.
But creating and storing a separate list of personal memories that a chatbot can bring up in conversations may raise privacy concerns. The company said what it was doing was not much different from the way search engines and browsers store their users' Internet history.