An Agent Can Remember Perfectly and Still Be Wrong: What Evidence Proves 'Done' After a Context Reset?

jylusdev · reddit · 2026-10-18

Core argument

The author highlights a common pitfall for long-lived agents: a checkpoint entry reading "deployment completed" is just the previous model's claim, not established fact. The next session reads it perfectly and continues — but nothing ever verified the deployment actually happened.

Typical failure case

Checkpoint says "done", but the deployment API still reports "in progress"
Health check shows the previous version
The execution response was lost when the session stopped

These records describe different things: what the agent claimed, what the deployment system reports, and what is currently observable. Flattening them into one phrase makes the handoff smaller while making the next decision worse.

Evidence worth preserving

The intended operation and its target ID
The execution receipt, if one exists
The observed target state and observation time
Any disagreement or missing evidence

The author also notes that a missing receipt shouldn't automatically trigger a retry — the action may have succeeded even though its response was lost.

Context

The author is building Jylus, which prepares source-backed context from records, relationships, and changing state, so the next model gets inspectable evidence instead of treating an earlier model's summary as authority. He closes by asking developers running long-lived agents: after a restart, what must exist before a task stays marked "done" — a target-system read, a durable execution receipt, or a task-specific check?

Original post →

An Agent Can Remember Perfectly and Still Be Wrong: What Evidence Proves 'Done' After a Context Reset?

More from AGI Musings