Durable agent runs — checkpoint, resume, audit

A crashed agent resumes from its last checkpoint, not from scratch.

error_outline

Without memory

Long-running agents — report generators, data pipelines, ops bots — die to timeouts, deploys, and crashes. Without durable state, every failure means restarting a multi-hour run from zero, and when something goes wrong you have no record of what the agent was thinking.

check_circle

With Rekall

Execution memory checkpoints every run. When an agent stops mid-task — crash, deploy, or timeout — the run resumes from its last checkpoint instead of from zero; every thought and step is recorded and streamable live, so runs are auditable during and after.

How it works

Runs checkpoint as they go

Agents persist execution state at each meaningful step — the plan, the cursor, the intermediate results.

Failures become pauses

A crash, timeout, or redeploy leaves a resumable run, not a smoking crater. Resume picks up from the last checkpoint with context intact.

Everything is observable

The run timeline and thought stream are live over SSE and stored for audit — what the agent did, and why, step by step.

See it in action

Checkpoint, die, resume

import Rekall from '@rekall/agent-sdk'
const rekall = new Rekall({ apiKey: 'rk_...' })

const run = await rekall.execution.create({
  task: 'Backfill 2M rows into the new ledger schema',
})
await rekall.execution.checkpoint(run.id, {
  step: 4, cursor: 'row:812000',
})

// deploy happens, process dies …

await rekall.execution.resume(run.id)
// → continues from step 4, row 812,000