The flight recorder, replayed

What did the agent know — and when did it know it?

When an agent makes a call mid-incident, the only fair way to judge that call is against the memory it had at that moment — not the memory it has now. BlackBox answers this with the database itself: the same consolidated recall the agent runs live is re-run inside an AS OF SYSTEM TIME transaction, and diffed against the present. No backups. No separate audit store. The memory is its own forensic record.

The recall to replay
2h ago

one embedding, two consistent reads — the historical one runs inside AS OF SYSTEM TIME

Why this is hard anywhere else

Point-in-time recall is a database property.

Consistent, not reconstructed

A snapshot, not a guess

The historical read is a consistent transaction at that timestamp — episodes, runbooks, and totals all from the same instant. Replaying from application logs reconstructs an approximation; this reads the actual past.

Same code path

The recall is the recall

The replay runs the live system’s own consolidation and hygiene-aware ranking — distance discounted by confidence — so “what it knew then” is computed exactly the way the agent computed it.

Audit answers

Blame the memory, or clear it

“The agent missed the fix” and “the fix had not been learned yet” are different verdicts. The diff shows lessons learned since, recurrence counts rising, and confidence earned or decayed — with timestamps.