Engineer builds self-updating runbook that learns from past incidents using AI memory

A software engineer grew frustrated with static runbooks that quickly became outdated and built a system that updates itself based on real incident history. The tool, built using a service called Hindsight, stores each incident as two separate memory entries — one for diagnosis and one for resolution — making past failures easier to retrieve during future alerts. A key feature explicitly records failed remediation attempts so on-call engineers are not misled into repeating steps that made problems worse. The system surfaces responses side by side, comparing answers generated with and without historical memory, so users can judge its value directly. The retrieval model combines semantic, keyword, entity, and temporal signals to match new alerts to relevant past incidents even when the phrasing differs significantly.