Mental-Imagery based Brain-Computer Interfaces (MI-BCIs) enable users to control applications using their brain activity alone, by realising mental-imagery tasks. Although promising, MI-BCIs remain barely used outside laboratories, notably due to the difficulties users encounter when attempting to control them. We claim that understanding and improving the user-training process could greatly improve users' MI-BCI control abilities. Yet, to better understand the training process, we need a model of the factors impacting MI-BCI performance. In other words, we need to understand which traits and states impact MI-BCI performance, how these factors interact and how to influence them to improve this performance. Such a model would enable us to design adapted and adaptive training protocols, to guide neurophysiological analyses or design informed classi-fiers, among others. In this paper we propose a theoretical model of MI-BCI tasks, which is the first step towards the design of this full cognitive and computational model.