We extend a previously developed model of routine action selection by incorporating functional components to support behaviour in a simple non-routine task sorting cards according to a rule that must be discovered by the subject. A minimal extension to the previous model, consisting of an activation-based working memory/inference system in which evidence is incorporated by simply exciting or inhibiting relevant rule nodes, is demonstrated to be capable of capturing basic performance on the task. The task is commonly used in assessing frontal brain injury, and the extended model is further shown to be capable of capturing the gross behavioural characteristics of frontal patients. However, it is argued that a purely activation-based working memory cannot capture the requirements of more complex tasks. The paper thereby demonstrates 1) how the basic routine action model might be extended to more complex behaviours, but 2) that such behaviours require more than simple activation-based memory processes to structure non-routine behaviour over time.