Filesystems for AI Agents: A Survey of an Emerging Landscape

AI agents — autonomous systems powered by large language models — are increasingly interacting with filesystems. They read and write files, navigate directories, use grep and glob for context retrieval, and rely on persistent storage for memory across sessions. What is surprising is how many of these agent-facing storage systems are rediscovering classic filesystem concepts: copy-on-write overlays, whiteout entries, inode tables, append-only journals, mount namespace isolation, and virtual filesystem layers routing paths to heterogeneous backends.

This talk surveys the rapidly evolving landscape of filesystems designed for or used by AI agents. We examine emerging open-source projects building SQLite-backed FUSE/NFS-mounted agent filesystems, virtual filesystem abstractions routing agent file operations to cloud storage and databases, MCP-based file access protocols, and the local storage architectures of coding agents. We connect these to recent academic work proposing OS-inspired agent architectures, semantic file systems, and file-system abstractions for context engineering.