This translates to "a mean of 3,751 correctable errors per DIMM per year": http://www.zdnet.com/article/dram-error-rates-nightmare-on-d...
I'm not sure how things pan out these days with newer memory types. ECC checks and fixes these errors so they're not an issue.
A far more recent study by CMU based on the entire fleet of Fb servers shows that correctable error rates dropped dramatically in the past decade.
http://repository.cmu.edu/cgi/viewcontent.cgi?article=1345&c...
Your statement though was interesting, how do you have both ECC memory report errors and 'undetected errors'. At least from a memory perspective, with ECC an 'undetected' error is a multi-bit error that both flips bits and leaves the ECC bits in a legal configuration. That seems like it would be pretty rare.
That said, I've seen motherboards (in our data center) where the memory slots themselves were unreliable (probably bad or weak solder joints on the DIMM sockets or missing terminator resistors). They appeared as a machine with a lot of ECC errors but the same DIMM in another motherboard gave no errors.
It did fall outside my expectation of how ECC works. One bit errors and three bit errors, but not two? Some access pattern that memtest strides don't hit? I didn't really need the extra RAM, so I just moved on without it.