This research presents the first systematic investigation into filesystem-based memory for Large Language Model (LLM) agents, treating a directory tree of Markdown files as a unified storage medium for declarative memory and procedural skills. The researchers formalized a framework involving three specialized roles: a management agent for content organization, a search agent for query answering with citations, and an execution agent for distilling task trajectories into skills. By evaluating various memory architectures—including agent-organized hierarchies, verbatim dumps, and chunk retrieval—the study tested how memory scale and tool harnesses (such as sandboxed shells versus specialized memory functions) impact performance. Findings indicate that organized memory structures significantly improve search economy, reducing retrieval costs by approximately 50% in large-scale datasets. However, the study reveals critical limitations in current agent capabilities: organizational integrity degrades over time for all but the strongest management agents, and none of the tested models demonstrated that improved organization directly translates to higher answer quality. Furthermore, the researchers discovered that the choice of toolset influences the memory store's structure as significantly as the underlying LLM itself. Ultimately, the study shifts the view of filesystem-based memory from a default assumption to a complex design space for future agentic architectures.