Reducing the memory in some sense always sacrifices accuracy, but of course if you are already using too high settings (e.g. too many k-points) you can still reduce the memory without getting worse results. Or, if you are putting several MPI processes on one machine, then don't; make sure each node only has one MPI process. In general, the memory usage is controlled by the number of atoms, the size of the basis set, the number of k-points, and the mesh cut-off. Reducing any of these reduces the memory usage. Otherwise, you just need to find a machine with more memory to run on