Developer fixes AI hallucination bug by moving counting logic out of the LLM

A developer building a customer support agent discovered that the underlying language model was producing inaccurate contact counts, sometimes undercounting or overcounting due to rephrased complaints. The root cause was that the prompt asked the model to both classify issues and count them — a task that blends judgment with deterministic arithmetic. The fix separated these responsibilities: the model now only classifies each interaction, while structured facts are stored in a memory layer called Hindsight and counted using standard code. This approach made the escalation logic auditable, since the final response includes both the computed count and the model's explanation for easy cross-checking. The developer noted that combining a vector-style retrieval store with structured fact storage eliminated the need to maintain two separate data systems in sync.