Moirai is an open format and reference implementation for recording, moving, and continuing AI agent sessions across tools. Agent work is usually locked inside one harness; Moirai preserves the useful history of that work, including events, checkpoints, forks, and artifacts, without requiring every agent to share the same runtime or expose its entire private context.
Ordered session events, immutable checkpoints and named milestones, forks with clear ancestry, and references to commits, patches, files, and other artifacts. Native checkpoints allow an exact resume in the original harness; portable checkpoints let another compatible harness continue the work.
Sessions can be private or public, with scoped collaborator access and a web interface for browsing history. Capture is explicit and inspectable, local storage is the default, remote storage and sharing are opt-in, and secret redaction happens before anything is persisted or exported. Moirai never records hidden chain-of-thought.
Moirai owns the portable session record; a harness decides what to capture and how to resume its own native state. October Bus can carry scoped references, but Moirai requires neither October Bus nor October. The Go reference implementation and TypeScript SDK are described in the Moirai repository.