In humans, working memory — our ability to hold and use information in everyday life — is closely linked to general intelligence.
That means the ability for AI to remember things could be the key to realizing a superintelligent AI, a still theoretical version of AI that reasons as well or better than humans.
OpenAI CEO Sam Altman thinks it’s hard to predict just how intelligent AI can really be because the possibilities of memory retention are limitless.
“Even if you have the world’s best personal assistant, they don’t, they can’t remember every word you’ve ever said in your life, they can’t have read every email, they can’t have read every document you’ve ever written, they can’t be looking at all your work every day and remembering every little detail, they can’t be a participant in your life to that degree. No human has like infinite, perfect memory,” Altman said recently on the “Big Technology” podcast.
AI, however, will definitely have the capacity for that, he said.
“Right now, memory is still very crude, very early,” he said. Once AI is able to remember every granular detail of a user’s life, including even the small preferences they didn’t explicitly indicate, it will be “super powerful,” he said.
Altman added that it’s one of the future features he’s most excited about — and he’s not the only one.
Andrew Pignanelli, the cofounder of The General Intelligence Company of New York, a company that builds AI agents for businesses, said that memory will become the biggest focus for AI companies in the coming year.
“It will become the most important topic discussed and recognized as the final step before AGI,” Pignanelli wrote in a blog post. “Every model provider will add and improve on memory for their apps after seeing OpenAI’s success with ChatGPT memory (like Claude just did).”
Pignanelli, however, said that the industry is still a long way from perfecting long-term memory.
“Larger context windows continue to improve things, as they allow more data to be passed into the context window, which allows the agent to better read parts of a large memory index,” he wrote, in reference to the amount of information a large language model can process in a single prompt. “Even then, though, the vast level of detail that we need to reach to consider something AGI requires memory architecture improvements.”
Even shorter-term episodic memory hasn’t been fully solved yet, he said.
Solving that memory problem is the ticket to turning AI from something that feels artificial to something that seems human, he said.
“Our systems today get the interaction part right. In terms of a Turing test for interaction, we’re basically all the way there. But that’s only half of what’s needed to make a digital self,” he wrote.
“The first AGI will be a very intelligent processor combined with a very good memory system,” he said.
