## TL;DR

The agent lost its checkpoint and restarted from zero. Persist checkpoints durably (database, not memory), and resume from the last one.

## Error

```text
lost checkpoint, reprocessed everything
```

## Steps

1. Persist the checkpoint after every batch to durable storage Expected: progress survives restarts.
2. On startup, load the last checkpoint Expected: the job resumes.
3. Skip already-processed records Expected: no duplicate work.
4. Verify a kill-and-resume cycle Expected: it works.

## When to use

- Checkpoint lost mid-run
- It was only in memory

## When not to use

- Checkpoints persist (then the resume logic is broken)
- The job is fast enough to rerun (then checkpoints are optional)

## Tool compatibility

- Checkpointing
- Long-running jobs

## Variant phrasings

### lost checkpoint reprocess

Persist checkpoints durably and resume from them.

### checkpoint in memory lost

Same: database, not memory.

## Why it happens

In-memory progress feels fine until the process dies at 90 percent. Agents learn this once, painfully.

## Edge cases

- Checkpoint the input offset and the output state together
- Corrupt checkpoints need validation on load

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_0SX_xYN8b5CEeQcQAoVVNw