AI Agent Hallucinating Facts? Stop Trusting Memory

LLMs are genuinely bad at reliably recalling specific facts — phone numbers, exact addresses, precise dates — from memory alone. They'll often generate something plausible-looking rather than admit they don't have the real value.

This is invisible in testing if you only ask about things the model happens to get right by chance. It shows up in production the moment a customer asks about something specific and gets a confidently wrong answer.

The fix:

never let an agent answer with a specific real-world fact from memory. Wire it to a real lookup — a database query, an API call, a document search — and have it retrieve the actual value instead of generating one. If the agent can't find the real data, it should say so, not guess.