What survives a context reset?

Video caption

The successor fills the gap with a plausible story.

Decision, fixture, trace and owning boundary.

A fresh agent must find and apply the constraint.

#EricFieldNotes

Full written post / accessible read

We give agents memory files and summaries because sessions end. That is sensible. But if the handoff says only 'retry handling fixed,' a future worker cannot tell which retry, which side effect or which test proved the fix.

In a simulated handoff, the old worker mentions a timeout patch but omits the provider ledger readback. The new worker assumes timeout means no charge, changes the retry key and confidently explains the wrong state transition. The failure is the missing evidence, not merely a short context window.

Attach the decision record to the source event and exact test. Keep a replayable fixture, a readback trace and the component that owns the invariant. Use retrieval to locate the packet; let a supervising role surface it when a related event appears, instead of flooding every worker with the whole archive.

Run a handoff drill: start a new agent on a nearby change, with no old chat. If it retrieves the packet, predicts the timeout case and keeps the protected test green, the memory path worked. Do this because a file existing on disk is not proof it reached the decision.

#EricFieldNotes

Evidence and boundary

On-screen label: ILLUSTRATIVE HANDOFF. Illustrative cases are not measured incidents. Research papers and vendor documents support the stated mechanism only within their studied or documented scope.