In humans, working memory (the ability to retain and use information in everyday life) is closely related to general intelligence.
This means that the ability of AI to remember things could be the key to achieving superintelligent AI, a theoretical version of AI that reasons as well or better than humans.
OpenAI CEO Sam Altman believes it is difficult to predict how intelligent AI will actually become because the possibilities for memory retention are limitless.
“Even if you had the best personal assistant in the world, they don't. They won't remember every word you say in your life. They won't read every email you write. They won't read every document you write. You can't go through every job every day and remember every detail, and you can't be that involved in your life. No one has a memory as limitless and perfect as humans,” Altman said recently on the Big Tech podcast.
But AI definitely has the ability to make it happen, he says.
“At this point, the memory is still very immature and very early,” he said. If AI can remember every detail of a user's life, including small preferences that the user didn't explicitly indicate, AI “will become very powerful,” he said.
Altman added that this is one of the future features he's most looking forward to, but he's not alone in it.
Andrew Pignanelli, co-founder of New York's General Intelligence Company, which builds AI agents for enterprises, said memory will be a big focus for AI companies next year.
“This will be the most important topic discussed and recognized as the final step before AGI,” Pignanelli said in a blog post. “Model providers will see OpenAI's success with ChatGPT memory and add and improve memory for their apps (just like Claude did).”
But Pignanelli said the industry still has a long way to go to perfect long-term memory.
Regarding the amount of information that large language models can process in a single prompt, he writes, “The situation continues to improve as the context window gets larger, as more data can be passed to the context window and agents can better read parts of large memory indexes.” “Still, thinking about AGI requires reaching an enormous level of detail and requires improvements in memory architecture.”
Even short-term episodic memory has not yet been fully resolved.
Solving this memory problem, he said, is the ticket to transforming AI from artificial to human-like.
“Our systems today get the interaction part right. As far as the Turing test of interaction goes, we're basically there. But that's only half of what it takes to create a digital self,” he wrote.
“The first AGI will be a very intelligent processor combined with a very good memory system,” he said.
