Ok Monkey, here's the thing.
If you are failing in memtest then you are failing in windows. However, the computer uses a technology called "ECC" which is Error Correction Code
. This means that when your ram is failing on a read/write/verify step it redoes the step until the verification is correct. If it finds a correct verify the code then passes the strap back to the program to continue running.
After five iterations of the error there is a flush of the cache and a new read/write/verify of the data occurs. This is done three times and if the failure continues to occur an errata event is created in the event log and in Prime95. Thus if your system is finding error (and it is if memtest is giving errors) it is validating the errored data through ECC validation and Prime95 is continuing.
Your system is still not stable although the ECC will allow you to continue your programs. Your ram is not creating stop events through non-valid ECC iterations.
I personally would drop your ram to a divider until memtest proves stable.
Here's the thing, every time your ram creates errors it is stalling the system. If it errors through to three or four iterations every five data streams of read/write/verify then your system is slowing down and all the overclocking you are doing is not meeting it's best overclock. It is the antithesis of overclocking.
Some will complain that their overclock seems slower at a certain speed. This is often because errata is being streamed and dealt with through ECC.
Hope this helps and enjoy your overclocking