LayerFS forks a writable repo per AI agent without copying files

LayerFS, a new open-source project from Ephemeral-AI-Lab, gives every AI agent its own writable copy of a project in milliseconds without copying a single file.

The repo's pitch is specific. In one documented run, an agent building an e-commerce site produced a commit history that added up to '770 MB of files, stored in a 6.3 MB database — 121× smaller.' The savings come from four mechanisms stacked together: content-addressed storage, content-defined chunking, copy-on-write trees, and delta encoding. Agents can fork from main, another branch, or a past commit at any depth with no copy overhead.

The tagline, in the project's own words, is 'Ephemeral Workspaces. Durable Shared History.' One workspace per agent or tool call. Commits merge rather than overwrite. Every line carries author attribution.

The caveats are stamped on the repo itself. Current version is v0.1.7, labelled preview. LayerFS 1.0 is scheduled for October 21, 2026, and until then the main branch is under active development with 'no guaranteed stability until official release.' Requirements are Linux on ext4, xfs or btrfs, plus Docker and Node 20+. No native macOS path.

Two of the researchers we track were already sharing the repo link soon after it went up.