Sandra knows what the latency values are set to in the bios, but the values themselves barely affect the results of the bandwidth test (as it should be, bandwidth and latency are two seperate things).
I don't know the exact details of the memtest's methodology, but it is extremely impacted by latency. This indicates the methodology is very different from sandra's, so no difference in its behavior comes as any surprise. Perhaps it was wrong to lump it in with a test like PCMark, as it is a different kind of test, but the important point is that its behavior falls much more closely in line with PCMark than it does sandra.
Bandwidth is defined as the amount of data that can be transferred per unit time, once a transfer begins. Latency is the time that elapses before an activity takes places, such as a data transfer. The overall effectiveness of a memory subsystem is the composite of these and other factors.
Just like trying to quantify hard drive performance with a single number, it is virtually impossible to do the same with memory. For some applications bandwidth is the dominant factor, for others latency is, some are affected strongly by both, some by neither. Just like drive performance memory subsystem effectiveness can vary all over the map depending on the exact task we are talking about, and each task is generally different. There is no way to simulate each and every one of the essentially infinite possibilities for the purpose of creating a composite score.
That being said, PCMark's memory test is the closest I have seen to a one-stop-shopping answer to memory subsystem performance evaluation. No benchmark is perfect, escpecially a low level syntetic one. But the results in PCMark are effected to a fairly realistic degree by both latency and bandwidth, and the results correspond well to the impact noted in real applications that are sensitive to memory subsytem performance (although many are not).
Sandra is a very useful comparison and diagnostic tool, but it is not a good benchmark. Its view of memory performance is too myopic to reveal much about the consequences of differing memory subsystem variables on realized application performance. It does however do a good job of isolating simple bandwidth from other factors, unfortunately most users are not in a position to put this one component of memory performance in its proper context.