File System as a Substrate for Memory and Cognition in AI Agents

This article puts forth the case that the operating system’s file system can be an effective cognitive layer for LLM-based AI agents. We draw from Unix, memory hierarchy, recent agent-file-system work and results from agent-memory benchmarks in explaining why the file system is a natural interface for LLM-based agents. We present a reference architecture in which identity, memory, skills, logs, and work products are stored in a transparent folder tree, and we describe results from local, self-hosted deployments for enterprise-level document management and personal AI use cases. We also discuss limitations (scalability, search, concurrency, schema drift) and propose decoupling the file system interface offered to the agent from the underlying persistence mechanism in order to keep the auditability and portability of file systems while allowing for persistence to databases.