Yes, we haven’t gotten around to finishing the docs there yet. A quick rundown is MEM_DF is a new JK algorithm which is optimal for beefy NUMA nodes, but the integrals cannot be reused as with DISK_DF
These algorithms produce identical results, just one is more efficient than the other. This only effects JK builds and no other parts of the code. The only restriction that many FNOCC and OCC programs assume that these integrals are on disk and so we need to use DISK_DF with these algorithms for SCF. SCF_TYPE is purely in the regime of SCF as noted.
The way it is currently setup is that a user should really only pick “DF” and have the backend algorithm choose the correct underlying algorithm for you. Can you post your inputs that were broken, we can take a look and make sure that these work as this change should be, in general, transparent to users.