An individual can only engage in a single review for a specific card at any given moment. It's impossible to conduct multiple reviews under identical conditions without parallel universes. This irreproducibility makes precise measurement of memory states impossible.

If we overlook the differences in card content and treat cards with similar review histories as a group, the stability and difficulty we measure are just averages, failing to address the underlying variability.

Due to the diverse nature of the data, any memory model trained on review logs essentially estimates mean values. The patterns of change in these averages may not accurately reflect the underlying mechanisms of atomic memory changes. Instead, shifts in distribution may be the main factor driving changes in mean values.

This means there are two effects changing the memory state: a) memory consolidation and b) content filtering. Every review involves both of these effects.