This paper introduces the learning mechanism by which agents can identify, through experience, important messages in the context of inference in a specific situation. At first, agents may not be able to immediately read and process important messages because of inappropriate ratings, incomplete non-local information, or insufficient knowledge for coordinated actions. By analyzing the history of past inferences with other agents, however, they can identify which messages were really used. Agents then generate situation-specific rules for understanding important messages when a similar problem-solving context appears. This paper also gives an example for explaining how agents can generate the control rule.