Large language models make the same memory mistakes as humans

Large language models struggle with working memory in the same ways humans do. They get confused when too much information is presented, and they remember recent items better than older ones.

Everyone assumed that because large language models can access all past information, they wouldn't have human-like memory limits. This paper shows that even with perfect access, these models still get confused by too much information, just like people. This means that simply giving an AI more context won't solve its memory problems; the problem is how it processes that context.