I offer you a simple conjecture: long-term memory is a bug, not a feature. To learn is to extract key patterns from events. Once that is done, the memory of events has no further value.
In an ideal world, we would learn from the environment around us, but remain unburdened by memories of the past.
But in the real world, learning is imperfect, and long-term memory is a work-around – it is an ‘archive’ of raw footage for us to consult, should we ever need them.
A more capable, or further evolved, learning machine (versus present-day humans) would not need long-term memory at all.
I forgot what I was going to say 😀