In a Spaced repetition memory system, the central noun is the prompt—a task constructed to promote Retrieval practice or to produce the Testing effect. But I’m interested in exploring an alternative formulation in which ideas are the core noun, or at least exist alongside prompts. To put this another way: to put something into a memory system right now, you encode it into prompts and tasks. What if, instead, you could put the thing itself into the memory system? And then, when you show up for your daily review sessions, you’d be presented with activities which reinforce and deepen those things?
This notion is typically discussed in the context of automation: good memory prompts are hard to write; it takes too long; wouldn’t it be nice if you could make the computer do that work? But let’s leave efficiency and convenience aside for a moment; I think there are more central reasons to consider “the thing itself” as the natural primitive:
In its most general form, we’re talking about a system which can facilitate practice sessions by selecting the most appropriate ideas to practice and providing appropriate activities.
In the case of the Mnemonic medium, this might mean a static library of activities or activity templates provided by the author. In the case of a more general system, this task probably requires some kind of generative AI (i.e. Large language models); see Using machine learning to generate good spaced repetition prompts from explanatory text. Challenges abound there.
I guess I’m proposing something like:
One thing that’s good about prompts-as-tasks is that they have a relatively clear natural chunk size. Novices do have trouble learning what that chunk size should be, but I think one can get a sense of it relatively quickly. Not so for ideas as the natural primitive. How complex should the idea be? Hm.
I notice that in practice, when I look at, say, University Physics - Young and Freedman, and try to pull out “an idea” which might be converted into prompts, I’m pretty dissatisfied. For example, consider: “Two positive charges or two negative charges repel each other. A positive charge and a negative charge attract each other.” OK… this is part of the idea of charge. There are two kinds of charge. What distinguishes the two kinds is exactly this property. And, in fact, nothing else distinguishes them. So… what kinds of questions might one ask about this idea?
These questions are sort of fine, as a starting point… but they’re too isolated, I think. You want this idea to interact with the idea of “proton” and “electron” and “conductor” and so on, in increasingly complicated ways.
“Idea” doesn’t quite capture what I mean. Some things which I might want to bring into my daily sessions which don’t quite fall under the term “idea”:
Some alternative concept-words to consider:
The fact that I can’t find a unifying word here probably suggests trouble for the idea I’m thinking about. Maybe it doesn’t make sense to use the same word for a concept which can capture all these things, because the corresponding activities will differ enormously.
I guess the main claim is that I’m interested in investigating this proposition: right now, you encode stuff you want to bring into your practice sessions as questions/tasks/prompts; instead, maybe you should be writing down the thing itself.
To put this another way: to put something into a memory system right now, you encode it into prompts and tasks. What if, instead, you could put the thing itself into the memory system?
A goofy observation: John Anderson’s ACT-R models condition in terms of production rules. “If I see two intersecting lines where I know the angle here and I’m trying to find the angle there, I can do X.” They’re quite rote. I’ve been thinking recently that they’re probably the wrong level of abstraction, at least for the instructional designer. It seems unrealistic to imagine that we’re going to model hundreds of production rules for every subject. And possibly quite rigid too. Better, perhaps, to develop some way to infer them from a higher-level definition. FORTRAN for production rules. I’m sure there’s been a lot of work done like that in the Intelligent tutoring system space; I’m not well-read in the research there.
Anyway, I notice that my complaints about memory systems above rhyme with my observations about ACT-R. A given prompt usually reinforces a narrow, ACT-R-like production rule. Just as the instructional designer really shouldn’t have to encode all the adjacent variations of production rules in ACT-R-based ITSes, a memory system user really shouldn’t have to worry about writing prompts which produce traces of all those adjacent rules. In both cases you want a higher-level language.