Well, your second screenshot shows that the signal on channel 1 has four different falling edges to choose from. Why shouldn't the scope "randomly" select any one of them to trigger, depending on which one comes along next when the scope is ready to trigger again?
I assume that, by setting the memory depth to 300k, you have influenced the scope's processing speed for each scan, and (by coincidence) have brought the scope into sync with the signal on channel 1. So the scope happens to always trigger on the same phase of channel 1 due to this synchronization.
If that's what is going on, the clean way for obtaining a stable trigger would be to use one of the more complex trigger conditions the scope offers. E.g. trigger on a falling edge preceded by at least 1µs "high" state or such, to always trigger on the first falling edge in the group of four.