Myoso, the final piece of the spaced-repetition puzzle.

For simple definitions and concepts, flashcards with active recall are perfect. But solving a problem? 100 problems, 1000 problems? It is not usual to have a system of spaced repetition with active recall in this case for some odd reason. Now there are attempts out there. I've seen many, the problem is they all fall for the same trap. An ordinary question prompt which suddenly leads into a heavy back-card with the entire solution squished into one place, or even a picture of the solution solved by hand.

This does not work. More specifically, it does not scale.

Simply because you are forcing too much mental load into one card. The purpose of flashcards is that they are atomic, and recall should be quick. If you go beyond a certain mental weight and you have a deck of 1000 problems to complete, you will never reliably reach high recall accuracy. It's very easy to blame the person, and claim they aren't working hard enough. But the reality is that this system does not scale efficiently, even more-so when the student isn't even using flashcards for their problem-solving at all.

I looked endlessly looking for some sort of system that has this all figured out. Endless wikipedia entries, blogs, papers. Nobody is really discussing this at all. I did come across a paper under the psychology of learning & teaching, flashcards-plus. This is a study which demonstrates that you can improve retention by formatting flashcards into individual steps. If there is a concept you want to memorise, rather than writing its definition in a flashcard, the user instead splits it apart into three components:

1. textbook definition
2. you're own words
3. real-world example or analogy

Doing this reliably increases recall accuracy very well. It makes sense why it worked, adding extra information in relation to the data you want to retrieve builds strong memory traces. This is loosely related to the idea of contextual retrieval in AI systems where adding additional related context next to the thing you want to remember, improves the accuracy of remembering that thing, on top of understanding it better in context. It's also why language-learners have found out that, putting vocabulary next to a real-world sentence in a flashcard, not only helps you understand the word better in context, but even improves recall accuracy of the world alone.

But here's the problem again, as soon as you start adding too much information to the back of a flashcard, you are now increasing its weight. Now if you read my older post, you'll know that this extra weight doesn't really matter when you are simply just re-reading your notes, however it becomes problematic when you are practising active-recall or the testing effect.

That's where I came up with the idea to extend ordinary flashcards, by instead dividing the mental weight of a heavy back-card into multiple individual steps. Your back-card is now n number of steps that build a long chain towards the solution, each step is individually rated as its own isolated card. So solving a hard problem can be either 2, 3 or even 10 steps if you wanted it to be.

The algorithm doesn't matter, you could use sm-2, leitner, ebisu. Myoso uses FSRS. However, the algorithm would need to be modified for for this multi-step approach to work. The user has to "unlock" each step as they go along, rating a step well means unlocking the next step. Scheduling the problem would involve taking into account the rating of each individual step towards the solution, and breaking an earlier step should require rebuilding the steps again. With all of these things into place, what you have is a system where you can encapsulate a problem of any size into a single flashcard without mentally exhausting the user during review.

The core piece is uninteresting on its own, each step is scheduled with FSRS, the same algorithm behind current-day Anki. Every step basically tracks its own individual stability (how many days until you're 90% likely to still recall it) and difficulty. So if you rate a step well, its stability obviously grows, so it comes back later. Rate it poorly and stability shrinks, so it comes back sooner. This part is essentially what every SRS implementation looks like behind the scenes, even older systems like Leitner and sm-2 are loosely doing the same thing, but not at the same level of sophistication.

The interesting part of Myoso is what happens around that per-step scheduling.