Stay informed with free updates
Simply sign up to the Artificial intelligence myFT Digest — delivered directly to your inbox.
Artificial intelligence companies are grappling with a challenge humans have encountered for as long as they have existed: how to retain memories.
OpenAI, Google, Meta and Microsoft have stepped up their focus on memory in recent months, releasing upgrades that enable their chatbots to store a larger amount of user information to personalise their responses.
The move is viewed as an important step to help top AI groups attract customers in a competitive market for chatbots and agents, as well as a means to generate revenues from the cutting-edge technology.
But critics have warned the development could also be used to exploit users for the commercial benefit, as well as raising privacy concerns.
“If you have an agent that really knows you, because it has kept this memory of conversations, it makes the whole service more sticky, so that once you’ve signed on to using [one product] you will never go to another one,” said Pattie Maes, a professor at MIT’s media lab and specialist in human interaction with AI.
AI chatbots such as Google’s Gemini and OpenAI’s ChatGPT have made huge strides. Improvements include expanding context windows, which determine how much of a conversation a chatbot can remember at once, and using techniques such as retrieval-augmented generation, which identifies relevant context from external data.
AI groups have also boosted the long-term memory of AI models by storing user profiles and preferences to provide more useful and personalised responses.
For example, a chatbot may remember whether a user is a vegetarian and respond accordingly when providing restaurant recommendations or recipes.
In March, Google expanded Gemini’s memory to a user’s search history — as long as the person gives permission — compared with previously being limited to conversations with the chatbot, and plans to expand this to other Google apps in the future.
“Just like with a human assistant . . . the more that they understand your goals and who you are and what you are about, the better the help that they will be able to provide,” said Michael Siliski, a senior director of product management at Google DeepMind.
OpenAI’s ChatGPT and Meta’s chatbot in WhatsApp and Messenger can reference past conversations, instead of just the current session. Users can delete specific memories from the settings, and are informed when the model is creating a memory in a message on the screen.
“Memory helps ChatGPT become more useful over time by making responses more relevant,” said OpenAI. “You’re always in control — you can ask ChatGPT what it remembers about you, make changes to saved memories and past conversations, or turn memory off at any time.”
For businesses, Microsoft have used organisational data to inform memory, such as emails, calendars and intranet files.
Last month, the tech giant started rolling out a preview of Recall on some devices, a feature that records user activity by taking screenshots of their computer screen. Users can opt out or pause the screenshots.
When it was first announced last May, it drew concern from the cyber security community and others who called it “creepy”, causing Microsoft to delay the launch several times.
AI companies are also betting that improved memory could play a big part in boosting monetisation through affiliate marketing and advertising.
Meta chief executive Mark Zuckerberg said on its earnings call last month that “there will be a large opportunity to show product recommendations or ads” on its chatbot.
Last month, OpenAI improved its shopping offerings in ChatGPT to better display products and reviews. The company told the Financial Times it did not have affiliate links “at this time”.
But the roll out of greater memory capabilities across LLMs has also raised privacy concerns, as regulators around the world watch closely how models may manipulate users for profit.
Increased memory can also lead to models trying too hard to cater their answers to users’ preferences, reinforcing biases or errors. Last month, OpenAI apologised after its GPT-4o model was found to be overly flattering and agreeable and rolled the model back to a previous version.
More generally, AI models can hallucinate, creating untrue or nonsensical responses, and experience “memory drift”, where memories become outdated or contradict each other, impacting accuracy.
“The more a system knows about you, the more it can be used for negative purposes to either make you buy stuff or convince you of particular beliefs,” said Maes, the MIT professor. “So you have to start thinking about the underlying incentives of the companies that offer these services.”
