An AI SRE agent needs two distinct knowledge types: the live plane (current metrics, logs, events) and the organizational plane (runbooks, postmortems, architecture docs, service ownership). The author's first build failed by stuffing organizational knowledge into the system prompt, which grew too large and became perpetually stale. The fix treats organizational knowledge as a retrieval problem: all runbooks, postmortems, architecture notes, and ownership maps live as markdown in a git repository. A pipeline chunks documents, embeds them, and loads them into a vector database. On every merge to main, embeddings for changed files are rebuilt, so the knowledge base is never more than one merge behind reality. At incident time, the agent retrieves only what the incident requires.