The complexity of Multi-Agent Systems is constantly increasing. With the growth of the number of agents, interactions between them draw complex and huge conversations. In this paper, we present a knowledge discovery process based on conversation mining and used to infer conversation models. Conversations are made up of sequences of messages exchanged inside the system. In order to discover the underlying conversation models that agents use while interacting with each other, we apply stochastic grammatical inference techniques on the conversations. We present some experiments that show the process we design is a good candidate to equip agents with the capacity to learn to interact.