Close your laptop lid and open it 1,000 years later to carry on your Baldur's Gate 3 save as if nothing had happened.
I once speculated to a friend about 15 years ago that eventually solid state storage space would be so fast that it could serve as active memory. I can't wait to tell him.
When I die and my laptop gets put into a box and shoved in a dusty corner forgotten for generations I want my descendant to be able to open it in the year 3000 and see all the tabs I had open when I died.
Such retention times are often probabilistic (ie. some percentage of bits have retained their proper value) and an "up to" value, which is negatively influenced by such things as the storage temperature and background radiation.
In practice, it might be that the only useful retention time is only a small digit number of years.
It should be fine for normal use cases when used with error correcting codes without any active scrubbing.
According error rates for ECC RAM (which should be at least by an order of magnitude comparable) of 1 bit error per gigabyte of RAM per 1.8 hours1, we would assume ~5000 errors in a year. The average likelyhood of hitting an already affected byte is approx. (5000/2)/1e9=2e-6. So that probability * 5000 errors is about a 1.2 percent chance that two errors occur in one byte after a year. It grows exponentially once you start going a past a year.
But in total, I would say that standard error correcting codes should be sufficient to catch all errors, even if in hibernation for a whole year.
I took this article specifically to mean, and that it was referring to, a new form of non-volatile solid state storage. Active memory is by definition, volatile. This article seems to be talking about non volatile RAM, fast enough to function as active RAM. This alone would redefine what a reboot is.
My initial thought was that everything would be stored in triplicate, then read in triplicate and 'voted' to the correct value, but I guess even that only extends the time before random bit-flips make the data unreadable. You're probably right on the need for active error checking if there is an intention to store anything long-term in this manner.
TMR (so the tripilicate method) wouldn't be super suitable for this kind of application since it is a bit overkill in terms of redundancy. Just from an information theory perspective, you should only have enough parity suitable for the amount of corruption you are expecting (in this case, not a lot, maybe a handful of bits after a year or two). TMR is optimal for when you are expecting the whole result to be wrong or right, not just corrupted. ECC and periodic scrubbing should be suitable for this. That is what is done by space-grade processors and RAM.