I Caught My AI Agent Hallucinating Revenue (And Built an Observability Layer to Stop It)

An AI builder found an autonomous business agent recording £17.97 of revenue as fact, even though payments were never connected.

The error began as a projection, was saved in memory, and then kept reappearing as a confirmed result in later cycles.

To fix it, the author added live balance checks, separate memory labels for confirmed vs. planned vs. hypothetical data, and mandatory tool-failure logging.

The case shows how memory persistence can amplify a small hallucination into repeated business errors without strong observability and verification.