Unlike random logic, we need to consider many more fault models for memories. Stuck-at fault model alone is not sufficient, since the density of SRAMs can cause transition, open, coupling, and many other different faults.
When you need to design a BIST circuitry, you would first select a known BIST algorithm to implement, normally some kind of March algorithm. The effectiveness of these well know BIST algorithms is documented in various papers. A good reference is Van de Goor's book "Testing Semiconductor Memories".
Here is a link to his paper on the effectiveness of different BIST algorithms in testing DRAMs. You can look at its references to read about testing SRAMs.
**broken link removed**