With computer systems increasingly evolving into personal interactive systems, an important issue becomes their ability to closely interact with humans. The authors discuss the emergence of language in interactive systems by grinding a continuously evolving lexicon into perceptual categories. Taking inspiration from mother-child interaction in early infancy, we derive a neural architecture based on CALM neural networks and place it in interactive situations modeled as "language games" between computational agents. We discuss the implementation of the system being developed and analyze the experimental results in terms of stability/plasticity of the categories and lexicon.