This paper proposes an agent model of case based classification. The idea is to allow cases, and more generally memorized problem solving experiences, to take a more active role in future problem solving. This is achieved through the so called memory agents which are selected cases with their own reasoning mechanisms. The proposed model of memory agents enables context sensitive classification, leading to a better overall classification accuracy. Memory agents enable localized incremental learning. The resulting system is therefore able to cope with the dynamic environment where class distribution may change over time. Evolutionary agent learning outperforms linear search and simple re-enforcement learning methods. Further, the agent model of memory indicates an important direction of study for memory based reasoning. It focuses attention on cases as agents with knowledge and intention rather than something passively waiting to be retrieved. The agent model provides a more suitable vehicle for integrating different AI techniques to form hybrid systems where different parts of the problem are better suited by different kinds of reasoning method.