How to Test Your RAM for Errors When a PC Crashes
For about six weeks my main machine crashed every three or four days. Never under load, which was the confusing part. Twice while a browser sat idle, once during a file copy, once at the Windows desktop with nothing running.

For about six weeks my main machine crashed every three or four days. Never under load, which was the confusing part. Twice while a browser sat idle, once during a file copy, once at the Windows desktop with nothing running. I reinstalled the graphics driver, then the chipset driver, then Windows itself, and the crash came back nine days later with a different stop code.
It was a memory error on one stick in the second slot, and it had been there since the day I built the machine. The reason it took six weeks is that I ran the built in Windows memory test early on, it passed in twelve minutes, and I crossed memory off the list. That test passing means much less than most people assume, and crossing an item off a list on weak evidence is how a fortnight of debugging becomes six.
What a memory fault looks like from outside the case
Memory errors are easy to recognise once you know the shape, because the shape is inconsistency. A failing component usually fails the same way. Bad memory fails differently every time, because what breaks depends on which program happened to be using the faulty address.
The signs that pointed at memory in my case, in hindsight: stop codes that changed between crashes rather than repeating, a zip archive that failed its integrity check and then extracted perfectly on the second attempt, a game that crashed to desktop with no error at all, and one saved file that opened with a block of garbage in the middle. Random is the keyword. If your machine crashes reliably in one game and nowhere else, that is a driver or a thermal problem and the shape is wrong for memory.
Things that look like memory and usually are not: a crash that only happens under heavy load, which points at the power supply or temperatures, and a machine that will not boot at all, which is more often a seating or storage problem. If you have not already checked temperatures while the system is working hard, do that first because it is quicker, and the method is in the notes on checking PC temperatures while gaming.
Which test to run, and for how long
The Windows Memory Diagnostic is not a serious test. It runs a short standard pass, it finishes fast enough to feel reassuring, and in my experience it catches only heavy faults. Mine passed it twice on a stick that a proper test failed inside nine minutes.
What you want is MemTest86 on a USB stick, run outside the operating system. It is free for home use, it writes a bootable drive in about two minutes, and running it outside Windows matters because it can then test the memory that Windows itself would have been occupying. Four full passes is the usual recommendation and that takes roughly two hours for every 16 GB. On my 32 GB machine a four pass run took just under eight hours, which is why I now start it before bed.
There is a faster alternative that runs inside Windows and loads memory harder in less time, which the overclocking community uses for tuning. For a machine that is simply crashing, I would still run the bootable test first, because an unstable system inside Windows can produce confusing results and you want a clean baseline.
One detail that caught me out. The default configuration runs every test in sequence, and the tests are not equally good at finding the common faults. Test 7 and the row hammer tests at the end are the ones that found mine. If you stop the run after an hour because the first few tests were clean, you have not tested the thing most likely to be broken.
A clean pass does not mean the memory is fine
This is where I think most advice on this subject goes wrong, and it cost me real time. Eight hours of clean passes is good evidence and it is not a verdict.
Two reasons. First, memory faults are often marginal rather than absolute, meaning a cell fails at a particular voltage and temperature and not otherwise. A test run from a cold boot in a cool room is the easiest possible condition. I now run an hour of a demanding game first, then reboot straight into the test, and that one change turned a clean run into a failing run on a friend's machine.
Second, and more common than genuinely faulty sticks, the fault is not the memory at all but the profile you are running it at. Enabling XMP or EXPO applies timings and voltages the memory vendor validated on their own test board, not on yours. The sticks are fine, the controller in your processor is near its limit, and the result is an error that looks exactly like a dead cell. In the last two years I have seen three machines where errors vanished completely after setting memory back to its stock JEDEC speed, and in each case the owner had been ready to send perfectly good sticks back.
So the useful question is not whether your memory is faulty. It is which of three things is true: a bad stick, a bad slot, or a profile your board cannot hold.
Narrowing it to one stick or one slot
Answering that takes a sequence rather than a single test. This is the order I use now, and each stage either finds the fault or rules out one of the three possibilities.
| Stage | How long | What a pass tells you |
|---|---|---|
| All sticks, memory profile enabled | 4 passes, overnight | Nothing is wrong at your daily settings |
| All sticks, stock JEDEC speed | 2 passes | Errors here mean hardware, not the profile |
| One stick, slot 2, stock speed | 2 passes | That stick and that slot are both good |
| Other stick, same slot 2 | 2 passes | Compares sticks with the slot held constant |
| Failing stick moved to slot 4 | 2 passes | Separates a bad stick from a bad slot |
| All sticks, profile on, after an hour of gaming | 2 passes | Your actual worst case is stable |
Note the fourth and fifth stages. Testing stick A in slot 2 and stick B in slot 4 tells you very little, because two variables moved at once. Keeping the slot fixed while you swap sticks, then keeping the stick fixed while you swap slots, is the part people skip and it is the only way to tell a dead cell from a dirty contact or a bent pin on the processor.
Before you start, write down what your memory is set to: the speed, the primary timings and the voltage. If a stage ends with you clearing the board settings, those numbers are gone, and I have watched someone spend an evening rediscovering a working configuration they had already found once. It belongs in the same place as the rest of your build notes, and the file I keep for that is described in documenting your PC setup.
What to do with a result is usually simple. One stick failing in two different slots means a return under warranty, and most memory carries a lifetime warranty so this is worth doing even on old parts. Errors that appear only with the profile enabled mean running at stock speed, which costs a few percent in games and nothing in anything else. Errors that follow the slot rather than the stick are a board or processor socket issue, and that is where I stop guessing and take it to someone with spare parts. Crashes that survive all of this are probably not memory at all, and the stuttering and crash checklist in the tips section covers the next things to look at.
FAQ
How many passes of MemTest86 are enough?
Four for a verdict, one if you only want to know whether something obvious is broken. A single pass catches a good share of faults but not the marginal ones, and marginal faults are exactly the kind that produce a crash every few days rather than every hour.
Can a memory error corrupt my files?
Yes, and it is the reason to take this seriously rather than live with the crashes. A bad address can land in data on its way to disk. If you have had a confirmed memory fault, check the integrity of anything important that was written during that period.
Is one error in eight hours still a failure?
Treat it as a failure. Memory is either error free across a full run or it is not trustworthy, and there is no useful threshold below which a reported error is acceptable. Rerun once to confirm it was not a fluke of a loose stick, then act on it.
Does the Windows test tell me anything at all?
A failure from it is meaningful and means you can stop testing. A pass means almost nothing. Use it as a two minute check when you want quick confirmation that something is broken, never as the reason to rule memory out.
More in Tips
Tips
How to Document Your PC Setup So a Rebuild Is Painless
My boot drive failed on a Friday evening in March. Windows reinstalled in about twenty-five minutes, which is the part...
Tips
How to Calibrate Monitor Colors Without a Colorimeter
A new monitor arrives, you plug it in, and the colors look wrong in a way that is hard to describe.
Tips
How to Back Up PC Game Saves Before Something Breaks
A reader wrote to me last spring after his boot drive died without warning.