No incident selected
Fire a demo alert to watch the agent recall what happened last time — and what it got wrong.
What the agent remembers
Every row is a real retained incident. Outcome, engineer and time are attached, so a fix can be audited — not just suggested.
The learning curve
A replay over the seeded corpus: at each incident, the agent may only use memories that existed before it. No future knowledge, no LLM.
Top-1 accuracy as memory grows
correct missed
Where memory helped
Where it did not — first time we had ever seen it
The incident corpus
Synthetic incidents across 4 services over 90 days, designed so failure classes recur and the agent can actually learn.
| Incident | Date | Service | What happened | Cause class | Outcome | Time |
|---|