One low-priority challenge for orchestration of Spaced repetition memory system sessions is Spaced repetition memory prompts should encode ideas from multiple angles, but often that means that practicing one prompt will “give away” the answer to another. The most trivial instances of this are “reverse” prompts (“The wave function is notated as…? $\Psi$; What does $\Psi$ denote in QM? The wave function.”) and prompts with multiple independent Cloze deletion regions. SRSs can sometimes special-case those instances. But it’s a general problem. Often it seems like it would really be best if a “cluster” of prompts about a single idea were always smeared out onto consecutive days, or at least separated by several minutes within a single session.
Some quick experiments with GPT-4 suggest that it can probably classify reasonably accurately whether two prompts “conflict” in this way. We could use this to “spread out” very similar prompts.