Skip to content
Checkpoint and resume

One atomic checkpoint. The same machine wakes.

A checkpoint is an atomic freeze of the whole micro-VM, memory, disk overlay and device state, stored as content-addressed chunks.

Scope and operating details

A checkpoint is an atomic freeze of the entire micro-VM: memory, disk overlay, and device state, stored as content-addressed chunks. Resume rebuilds that exact machine, reconciles the host and guest ledgers by sequence number, and the agent continues where it stopped.

Saved pointMemory and files
InterruptionThe running process stops
ResumeContinue from the checkpoint
Illustrative workflow. The evidence and limitations below define the tested scope.

What a checkpoint contains, literally.

Guest memory: the full RAM image of the micro-VM, including every process, open file handle, and the agent harness mid-thought

Disk overlay: every byte the session wrote: workspace edits, build artifacts, installed packages

Device state: the virtual machine state file: vCPUs, timers, virtio devices

The two ledgers: the host event log and the in-guest transcript, reconciled by sequence number before the agent wakes

What it deliberately does not contain.

The model provider's server-side state (prompt caches, KV caches) belongs to the provider and cannot be frozen from outside; a resumed session re-warms it on first use. Interrupted model CALLS are a separate mechanism: the gateway stores completed responses and replays them on retry, so the tokens already bought stay bought.

Corruption refuses to restore: a checkpoint with one flipped byte fails content addressing and no machine boots from it. A checkpoint that cannot be trusted is a checkpoint that does not run.

Where checkpoints live.

On the hosted fleet, checkpoints live in KeepState storage as content-addressed chunks, deduplicated before they are ever billed, and kept while your account exists with no automatic expiry, per the published retention and deletion policy. The storage meter charges the deduplicated footprint, and account-deletion tests are separate from customer per-session deletion. Deleting a session erases the content only it held, in production since 6 October 2026; stopping a session does not delete its saved content.

Checkpoint publish

crash-consistent MEASURED

chunks + fsync + temp manifest + atomic rename; a crash mid-publish leaves the previous checkpoint byte-identical

evidence
badge
MEASURED
measurement
crash-consistent · chunks + fsync + temp manifest + atomic rename; a crash mid-publish leaves the previous checkpoint byte-identical
test scope
Laboratory or bench evidence; production behavior is not inferred from it.
date
28 July 2026

Agent session resume

13.7-25.2 s MEASURED

measured (gates 4, 6, 7); about 4 ms of that is the hypervisor restore

Storage after dedup

19.1x MEASURED

one soak, 82 checkpoints: 4,233 MiB actual vs 80,952 MiB naive. One workload, stated as such.

evidence
badge
MEASURED
measurement
19.1x · one soak, 82 checkpoints: 4,233 MiB actual vs 80,952 MiB naive. One workload, stated as such.
test scope
Laboratory or bench evidence; production behavior is not inferred from it.
date
31 July 2026

Historical tests checked crash-consistent publication and rejection of corrupted saved state. Public summaries retain their dates, results and limits. record

Checkpoint your own workload.

Sign up and your sessions checkpoint on our fleet from the first second; the recorded kill shows exactly what that looks like.

keepstate.ai · the durable agent session cloud