For real? I always assumed a few passes would be enough! :| A quick google search seems to indicate that 1 night at most should be enough?
Cite? You have a quote of someone who is competent at building server gear, or is this just some random YouTuber who figured out how to bodge together a desktop and happened to know about memtest86?
Sorry if that sounds harsh, but the average use case for memtest86 is for someone to identify that bad or not-quite-compatible desktop RAM.
Have you seen it pass for a few cycles only to crash later on? I'm curious here :)
For real.
The burn-in time on a machine should exceed a thousand hours, which gives you a fair degree of certainty that you're past various infant mortality issues.
This is not your desktop machine. This is a machine that you need to be rock solid. Especially in the situation being discussed here, a machine that lacks ECC and therefore lacks any way to detect subtle problems after the fact, the only way to detect marginal memory is to test it repeatedly, hoping to catch that bad behaviour. Especially since the machine has been shown to corrupt data.
Is a week overkill? Perhaps. You're 98% likely to find a bad DIMM on the first pass. So if we're strictly talking memory testing, maybe a week isn't really buying you that much. But consider again what I said above:
When you look at the 5 year operational lifetime of a typical server, one week is 1/250th of that, so if it can go 1/250th of its life with no errors, the likelihood of an undetected error popping up later is much reduced.
Since the burn-in period is over a month anyways, there's no harm in running a week of memtest86 at the start and also at the end, and then run disk tests in between. That month isn't just for the memory. It's to allow disks to fail, for fans to show themselves as marginal, and for all other parts of the system to be strenuously tested and to fail.
A platform that you just throw together, run a one-pass memtest on, and then go ahead and load FreeNAS on and start putting your data on there, that seems to have about a 10% chance of ending badly.
A platform that you carefully put together and test for 1000+ hours, testing the way I've described in the burn-in document, that's a much safer place to house your data.
None of you have to listen to me. Bully for you if you don't! Hah.