At a fundamental level the algorithms predict the probability of a learner to correctly recollect a factoid at a given point in time given a history of sampling that recollection / presentation.

It would be interesting to have machine learning predict these probability evolutions instead. Simply recollecting tangential knowledge improves the recollection of a non-sampled factoid, which is hard to model in a strict sense, or perhaps easy for (undiscovered) dedicated analytic models. Having good performing but relatively opaque (high parameter counts) ML models could be helpful because we can treat the high parameter count ML model as surrogate humans for memory recollection experiments and try to find low parameter count models (analytic or ML) that adequately distill the learning patterns, without having to do costly human-hour experiments on actual human brains.