We present an experiment for the study of learning in a complex task which requires both memorisation and the ability to process several pieces of information. The outcome of an action, for which immediate feedback is given, depends on the context (i.e. one of thirty-two sequences of three features) which is know and visible to the subjects. Subjects develop some theories of the experimental world, which result in the stable repetition of some actions in response to certain conditions. These theories are modified after feedback, however mistaken answers are repeated and correct answers abandoned. During the game, theories become more effective (i.e. they afford more correct answers and a higher score), yet the improvements slow down. The theories follow from only a portion of the available information and when they become successful (i.e. towards the end of the experiment) the subjects start refining them to include a larger subset of the information, this causes more stable mistakes.