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

Researchers systematically evaluated filesystem-based memory for LLM agents.

Structured hierarchies sharply reduced retrieval cost for large memory stores, but the organization often decayed over time when the management agent was weak.

File organization alone did not reliably improve answer quality, and tooling choices could reshape memory structure as much as changing the model.

The study frames agent memory as a design space where efficiency, maintainability, and robustness must be balanced as stored knowledge grows.