The paper proposes a concept for a knowledge representation approach that consists of the knowledge structure in the form of ontology, knowledge base in the form of rules and environment model in the form of objects in the environment. The approach is developed for domains where the main problem is to choose the most appropriate capability for a particular action. Both the agents’ knowledge and also knowledge structures can be edited by the user after launching the system. An Ontology Learning Tool is implemented for this purpose. Additionally the knowledge base can be complemented by the agents themselves as a result of the learning process that is based on the user’s feedback. The proposed approach is explained in this paper on an example of vacuum cleaning multi-robot system. The implementation is done in the form that is suitable for development of JADE based multi-agent systems. The first validation of the concept is intended in a virtual multi-agent environment where JADE agents simulate cleaning robots.