This is a flaw of the fundamental nature of the service. If it were a clientside application, keeping track of your entire viewing history forever would be trivial. But because it's on a remote server, keeping everyones history forever would be very difficult. I would guess they're probably using a bloom filter and just not dedicating enough resources to it per-user. Or, maybe even more likely, they ran 1 experiment 1 time on a few users and discovered that re-watching previously watched content was popular and extrapolated from that - ignoring that their sample size included a great many children (since they outnumber adults) resulting in them just expecting everyone to act like children.