It has been recognized that AI programs suffer from a lack of generality, the first gross symptom being that a small variation to the problem being solved usually causes a major revision of the theory describing it. The lack of generality seems an unavoidable consequence of the process of approximating the world while building theories about it. In this paper we propose an approach where generality is achieved by formulating, for each problem at hand, an appropriate local theory, a theory containing the needed information. The process of theory formulation and reformulation is formalized using contexts.