Research prototypes in AR usually do not emphasize software architecture. Nevertheless, their architecture is not arbitrary, but results from specific needs. Architectural approaches embodying research contributions are of particular value for reuse at component and architectural levels. We have conducted a study of existing AR software architectures for the ARVIKA project [10]. This has resulted in a catalog of important desired quality attributes for AR systems, a reference architecture for comparison of AR architectures, and a catalog of architectural approaches used in current AR systems. We believe this lays the foundation for further research in AR software architectures.