If an AI is going to help you over weeks or months, it has to remember what you told it. There are two basic ways to pull that off, and people mix them up constantly. Agent memory and RAG both put old information back in front of the model. They just go about it differently.
And in October, a study led by researchers at the University of Illinois found the fancier approach often loses to the plain one.
Agent Memory
Agent memory is a system that turns what you’ve said into notes for later. It summarizes, reorganizes and decides what matters (”prefers short emails,” “uses Notion for everything”), then stores that. Next time, the agent pulls up its notes instead of the whole history. A lot of the “it remembers you” features in chat apps work roughly like this, at least partly.
It’s compact and quick. The catch is that summarizing means deciding what to throw away, and sometimes the thing it tossed is exactly what you need three weeks later.
RAG
RAG (retrieval-augmented generation) skips the note-taking. It keeps the raw history, and when you ask something, it searches that history and pastes the most relevant chunks into the model’s prompt. Nothing gets summarized, so nothing gets lost in a summary. It’s the same trick behind a lot of document search tools, just pointed at your own conversation history.
It’s less elegant, and it can grab the wrong chunks. But the original evidence is all still sitting there.
The tell
Ask yourself: is the AI reading its notes, or rereading the transcript? Notes means agent memory. Transcript means RAG. It’s the difference between an assistant coach’s notebook (”their #7 always cuts left”) and rewinding the game film.
The study puts a number on it. On one long-conversation test, plain RAG scored 90%, and the 11 memory systems it went up against scored between 63% and 84%. The researchers call the gap the “memory tax,” meaning the questions a memory system gets wrong that plain retrieval gets right.
Their fix was to use both: check the notes and search the raw history at the same time. That closed a lot of the gap. One system went from about 74% to about 83%.
It’s a preprint and mostly one benchmark, so I wouldn’t rip out anybody’s memory system over it. On a different test, a couple of the memory systems actually beat RAG. But it’s a decent reminder that a clever summary is still a summary, and sometimes you want the film.

