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

A recent study published on arXiv (2607.26637) investigates filesystem-based memory for LLM agents, questioning the dominant focus on custom memory structures in current research. The authors outline a framework involving three distinct roles within a centralized memory filesystem: a management agent that organizes incoming data, a search agent that responds to inquiries with referenced materials, and an execution agent that generates task trajectories distilled into skills. This method combines declarative memory and skills into a unified storage system, examining whether an agent can keep the organization intact as memories grow, conflict, and degrade. The research challenges two assumptions: an agent's ability to maintain an organized memory and the advantages of such organization. It marks the first comprehensive study of this standard memory medium, increasingly utilized by LLM agents through generic file tools for managing markdown files in a directory structure.