An approach to incremental classification learning is proposed. Classification learning is based on approximation of a given partitioning of objects into disjoint blocks in multivalued space of attributes. Good approximation is defined in the form of good maximally redundant classification test or good formal concept. A concept of classification context is introduced. Four situations of incremental modification of classification context are considered: adding and deleting objects and adding and deleting values of attributes. Algorithms of changing good concepts in these incremental situations are given and proven.