Filesystem-Based Memory for LLM Agents: Organization, Evolution, and Sustainability
Organized filesystem memory cut retrieval cost, but did not improve answer quality in the study.
The paper tests LLM-agent memory stored as directory trees of markdown files, with agents managing, searching, and distilling task experience into the same filesystem. Across benchmarks and embodied tasks, organization mainly helped search efficiency, roughly halving retrieval cost when the memory store was large. The authors found organization degraded as memory grew except under the strongest management agent, and measured agents did not turn better structure into better answers. Tool choices also changed the shape of the memory store about as much as model choice. HF Daily Papers' note
The paper tests LLM-agent memory stored as directory trees of markdown files, with agents managing, searching, and distilling task experience into the same filesystem. Across benchmarks and embodied tasks, organization mainly helped search efficiency, roughly halving retrieval cost when the memory store was large. The authors found organization degraded as memory grew except under the strongest management agent, and measured agents did not turn better structure into better answers. Tool choices also changed the shape of the memory store about as much as model choice. HF Daily Papers' note
score 5