Ontologies underpin the semantic web; they define the concepts and their relationships contained in a data source. An increasing number of ontologies are available on-line, but an ontology that combines information from many different sources can grow extremely large. As an ontology grows larger, more resources are required to use it, and its response time becomes slower. Thus, we present and evaluate an on-line approach that forgets fragments from an OWL ontology that are infrequently or no longer used, or are cheap to relearn, in terms of time and resources. In order to evaluate our approach, we situate it in a controlled simulation environment, RoboCup OWLRescue, which is an extension of the widely used RoboCup Rescue platform, which enables agents to build ontologies automatically based on the tasks they are required to perform. We benchmark our approach against other comparable techniques and show that agents using our approach spend less time forgetting concepts from their ontology, allowing them to spend more time deliberating their actions, to achieve a higher average score in the simulation environment.