It doesn't matter whether this actually models how memory works in reality, as long as it has actual predictive capabilities which correlate to how it works in practice.
And this can be easily verified with data and simulations. The algorithm predicts how likely it is that you'll remember a given piece of information (e.g. in three days you'll remember this card with 80% probability), so if you can get a big dataset of reviews and run it through the algorithm you can easily check how accurate it is, e.g by calculating a brier score, or by comparing predicted vs actual recall curves.
Source: I've developed an even better algorithm than FSRS (I've directly compared them in the past, although that was quite a while ago so it might have been for one of the older versions of FSRS, so I'm not sure how the newest one compares), and now I'm working on an even better one.
And this can be easily verified with data and simulations. The algorithm predicts how likely it is that you'll remember a given piece of information (e.g. in three days you'll remember this card with 80% probability), so if you can get a big dataset of reviews and run it through the algorithm you can easily check how accurate it is, e.g by calculating a brier score, or by comparing predicted vs actual recall curves.
Source: I've developed an even better algorithm than FSRS (I've directly compared them in the past, although that was quite a while ago so it might have been for one of the older versions of FSRS, so I'm not sure how the newest one compares), and now I'm working on an even better one.