Features · State Branching

Git for AI memory

Branch agent cognitive states to test alternative reasoning paths without corrupting main memory history. Fork, experiment, merge, or discard.

Instant snapshotting

Fork a memory space with `claw branch`. It's copy-on-write, so it costs almost nothing until it actually diverges.

Parallel reasoning trees

Run several speculative agent paths against isolated forks, then merge or discard each independently.

Clean merges back to trunk

Bring a successful branch's changes back into production memory without dragging along everything it tried and discarded.

Deterministic rollback

Replay or roll back a branch to any prior state without touching trunk memory at all.

Why branch memory at all

Multi-agent systems and speculative planning both need to try something without risking production state. A branch gives an agent, or a whole swarm, an isolated copy of memory to reason against, then a clean merge path back to trunk once the outcome is known to be good.

claw branch create --from trunk --name scenario-a
In practice

What this looks like day to day

A/B testing agent behavior

Run two versions of a prompt or policy against identical forked memory, compare outcomes, and merge whichever one actually performed better.

Sandboxing a risky operation

Let an agent attempt a multi-step task on a branch first. If something goes wrong, discard the branch and trunk memory was never at risk.

Parallel multi-agent exploration

Fan a research question out across several branches explored by different agents in parallel, then merge only the branch that found the answer.

FAQ

Common questions

No. Branches are copy-on-write, so creating one is close to instant regardless of how much memory is in trunk. Storage cost only grows as the branch actually diverges.

Give your AI agents permanent memory today.

Deploy in under 2 minutes. Free tier included forever.

Launch Studio Free →