It is simple: in UV erasable EPROMs (also one-time programmable versions, which use same internal technology), an 'empty' memory bit reads as a 1. And you can program a 0 into each bit. This is done by using the programming voltage (25/21/12~12.5V) to pull a bunch of electrons onto an isolated 'island' (the floating gate). Erasable by shining UV light through the window to kick those electrons off in under a half hour (a few weeks in sunlight). For any proper functioning EPROM this floating gate will be
very well isolated from its surroundings. Read: able to keep that charge for many years.
But: no matter how well constructed, this isolation is never
perfect. There will always be ways that allow small numbers of those electrons to find their way out of their 'cage' (cosmic radiation, tiny amounts of UV in normal lighting that might pass through a label, tiny amounts of moisture sealed inside the IC package, etc, etc). Thus that floating gate is leaky, and the 0 it keeps
will flip back to a 1 if only you wait long enough.
That might take much longer than manufacturer said it should take. Or much shorter. Regardless of reason, 20 years is a long time, and for 20+ years ago programmed EPROMs it is not unheard of that some programmed bits flip back to erased state. Not "all the time", but it happens (I myself have seen it a few times). Simple as that.
Microchip wrote:Data Retention > 200 years
IC's buit today won't be functional anyway when they get 200 years old. So for all practical purposes, that is a bullshit claim. At best, it's an indicator of expected performance over shorter period like 10 or 20 years. So take a claim like that with a biiiig grain of salt...