Three LLM judges couldn't tell an AI agent's real memory from a counterfeit one