This paper investigates how an intelligent teaching agent with Role and Reference Grammar [RRG] (cf. Van Valin 2005) as linguistic engine can support language learning. Based on a user-centred empirical design study the architecture of a highly persuasive tool for language learning as an extension of PLOTLearner (http://europlot.blogspot.dk/2012/07/try-plotlearner-2.html) is developed. Based on grounded theory it is shown that feedback and support is of greatest importance even in self-directed computer assisted language learning. Is also shown how this overall approach to language learning can be situated into traditional conversation based learning theories (cf. Laurillard 2009). It is shown that a computationally adequate model of the RRG-linking algorithm, extended into a computational processing model, can account for communication between a learner and the software by employing conceptual graphs to represent mental states in the software agent and the important role of speech acts is emphasized in this context.