Intel / article

Filesystem-Based Memory for LLM Agents: Organization, Evolution, and Sustainability

Source: arxiv.org
Author: Sizhe Zhou, Sheldon Yu, Hui Wei, Junda Wu, Siru Ouyang, Yizhu Jiao, Shijia Pan, Julian McAuley, Yu Zhang, Tong Yu, Jiawei Han
Date: Published Jul 30, 2026

Context

Organizing an 's saved files made information cheaper to retrieve without making its answers better in this study. Researchers compared agent-organized folders, unorganized text and chunk retrieval across conversation and task benchmarks. Organized stores roughly halved retrieval cost when the material was large, but none of the tested agents turned organization itself into improved answer quality. As memories accumulated, organization also deteriorated for all but the strongest management agent. The result gives filesystem memory a specific benefit to measure: reducing search effort while separately checking whether retrieved information improves the answer. Study abstract.

Key takeaways · AI-distilled

The paper formalizes filesystem memory as three roles over one directory tree: a management agent that files and reconciles incoming content, a search agent that answers queries with citations, and an execution agent whose task traces are into skills.
What organization actually buys is search economy. An organized store roughly halves retrieval cost once the material is large. It does not buy better answers: no agent measured converted structure into higher answer quality.
Organization decays under growth. As the memory stream scaled, every management agent except the strongest let the store erode, with conflicts and stale entries accumulating faster than the agent tidied them.
The tool is as strong a lever as the model. Swapping the file tools alone, shell versus memory-tool-style functions versus different search tooling, reshaped the resulting store about as much as swapping the model did.
The comparison that makes it credible: agent-organized hierarchies are tested against a verbatim dump and plain chunk retrieval, so the industry default of telling an agent to organize its notes is measured against not bothering.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
distillation — Training a small, cheap model to imitate a big one's outputs, keeping much of the capability at a fraction of the cost.
agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
sandbox — An isolated environment where AI-generated code or agent actions run without being able to touch anything real.

Key quotes

“What organization reliably buys is search economy: organized stores roughly halve retrieval cost where material is large.”

“Today's agents, however, fall short of the default's promise: in our growth study, organization erodes for all but the strongest management agent, and no agent we measure converts organization itself into better answers.”

“And the model is not the only lever over a store's shape: changing the tool set alone reshapes the store as strongly as swapping the model.”

“The study turns the filesystem default from an assumption into a design space for agent memory.”

Read the source arxiv.org