The complexity of inference using the belief propagation algorithms increases exponentially with the maximum clique size. We describe an approximate inference approach when there are clique size limitations due to memory constraints using incremental construction of clique trees.