Study: filesystem memory halves LLM agent retrieval cost

Something quietly interesting from arxiv this week, because it points at where agent research is starting to look more like systems work than model work. A new paper from Sizhe Zhou and collaborators, posted to arxiv, presents what the authors call the first systematic study of what happens when you give an LLM agent a filesystem, a directory tree of markdown files, as its long-term memory, and let it read, write, and reorganize the tree over time.

The headline finding is a nice piece of engineering good news and a slightly awkward piece of research bad news at the same time. Organized stores, the authors report, "roughly halve retrieval cost where material is large." That is the kind of number that changes deployment economics for agent products that grind through big knowledge bases. But the second finding is the awkward one: "no agent we measure converts organization itself into better answers." Structure buys you speed and cost, not quality.

The other two results are the ones that will probably matter most for anyone building on this. First, "organization erodes for all but the strongest management agent," meaning that in practice the tidy directory tree turns messy as the agent lives longer, and only the best management-role model keeps it clean. Second, and less obvious, "changing the tool set alone reshapes the store as strongly as swapping the model." Tool design, in other words, is a first-class lever over how an agent's memory actually looks, not a lower-order implementation detail.

The honest caveat is that this is one paper on a young research question, and the framing is careful. It formalizes three agent roles, management, search, and execution, around a shared filesystem, but what the reporting does not give you is a prescription for which management-agent architecture actually sustains organization at scale, or which specific tool designs help versus hurt. Those are the follow-up experiments the field will want.

If you are shipping agent products, the practical read is straightforward. Filesystem memory looks like a real cost and latency lever if you are willing to invest in the management agent that keeps the tree tidy. Just do not sell it internally as an accuracy win, because the paper says it is not one yet.