Our group at SRI is addressing several limitations of FRSs by building on existing database systems. FRSs cannot support high-speed access to large knowledge bases, nor do they have multi-user access capabilities. We have designed a storage subsystem for FRSs by submerging an existing DBMS within an FRS. We report on empirical evaluations of several variants of this architecture. We compared the performance of different DBMSs (relational and object oriented). We also developed and evaluated several optimizations including transferring data at different granularities, and the use of prefetching. The resulting system is in use in our group to manage information for the human genome project.