9 min read

Built on the Science of Memory

noruditmethodologylearning sciencememory

The foundation of The Norudit Academy can be classified into 2 parts, the design and the method. In this post, we'll address the ideas that are the basis of the method that led to the creation of the Norudit Flywheel.

Most study apps are built around one or both of two angles: engagement farming, like streaks, XP, games, etc, or what they believe is productive, what I like to call productivity scams, like rereading, highlighting, summarisation, etc.

In this post, we'll focus on the productivity scams. If you'd like to learn more about engagement farming and Norudit design, check out the post on Norudit design.

Knowledge tracing: a live map of what you know

The first step to studying is knowing what to review/learn. Why? It tells you where to focus your attention, reducing the risk of spending time on areas that should have been deferred or skipped entirely.

A lot of tools get this wrong: they use baseless and often ineffective methods, like a progress bar, to track learning. A high percentage on a progress bar misleads the student into believing they understand the concept, but that's not how learning works, especially in the long term. More on that below.

Learning has its ups and downs: you may learn a concept today, and tomorrow it's fuzzy. That doesn't mean you don't know it, just that your mastery fluctuates.

That brings us to the first basis of the Norudit Flywheel: Bayesian Knowledge Tracing (BKT).

Bayesian Knowledge Tracing (BKT) was introduced in 1994 by Albert Corbett and John Anderson as a foundational statistical model to track how a student's understanding of a specific skill evolves over time. BKT is a predictive algorithm used primarily in Intelligent Tutoring Systems (ITS) to model whether a learner has mastered a specific concept. It treats a student's internal knowledge state as a "latent" (hidden) binary variable, meaning the system cannot see what the student knows directly, so it must infer it. The model updates these probabilities after every single interaction (such as getting a math problem right or wrong) using Bayes' theorem.

Or in layman's terms, it's an algorithm designed to keep track of a student's mastery of a concept, and in turn used to decide which topic to focus on next. A mathematical model like this ensures the right amount of effort goes into the right topic, instead of relying on "vibes" to determine your weak point.

This algorithm is also used in the mock exam generator in The Norudit Academy. You can read more on the foundational principles behind how Norudit builds questions for the mock exams and Phase 6 of the Norudit Sequence here.

Spaced repetition: reviews timed to forgetting

So we know what to review, now let's talk about when.

In 1885, Hermann Ebbinghaus measured his own forgetting and drew the curve that would become the foundation for modern spaced-repetition algorithms, from the older SM-2 algorithm to FSRS-6, the newer model now offered in popular apps like Anki (Anki previously used SM-2 but updated to FSRS-6 because FSRS offers significantly higher efficiency, reducing review workloads by 20-30% while maintaining the same target memory retention). The core idea is that memory decays fast at first, then levels off, and each time you successfully recall something, the next drop is slower.

A century of work turned that curve into a rule: the same hour of study spread across days beats the same hour crammed into one.

Cepeda et al., 2006

Fixed-card tools like Anki proved the science works. Norudit schedules with the same FSRS-6 algorithm, but Anki, although great, has a lot of flaws.

For example:

Concept-varied spaced repetition solves the problem where students memorise the card rather than testing their understanding of the concept over the long term. My solution was for the cards to vary while maintaining the underlying concept, with the degree of variation depending on the card's concept mastery. If you'd like to read more on concept-varied spaced repetition and how we fixed the other problems with Anki and spaced repetition tools in general, check out the deck.

Retrieval practice: recall it, don't reread it

Now on to the aspect I believe most tools genuinely get wrong, and a trap most people fall into, not just students. The core problem is human nature: we seek out comfort over tasks that require effort. That's why the most common techniques, like rereading and highlighting, tend to rank highest in usage, while ranking lowest in terms of utility, because they're "easy". Read more about the 10 most common techniques.

The issue is we live in a "soft era" where everyone is looking for the easy way out. But that's nothing new, this has been the pattern for millennia, which is quite ironic, since these easy methods tend to have the exact opposite effect. They're easy and feel productive, so people do them anyway, and when they fail, which is often, they blame society, their school, or themselves, saying things like "I'm just not good at [insert a skill or subject]." These popular tools capitalise on that and sell you the ease. We've seen this boom with the advent of AI too. You've probably heard the pitch: "AI will do the work for you, you just do nothing." People buy the idea of doing nothing to feel productive, then wonder why they can't improve.

So what's the solution? It's pretty simple, but not that clear for most people: do the work. So the question now is, what work should they do? The research shows the best technique is to practice recall: things like practice questions, the Feynman technique, and blurting. These techniques force you to recall a piece of information, and doing this frequently, even across different subjects, trains your brain to recall information faster and better.

The brain works like a muscle: the more often you use it, the stronger and sharper it becomes.

And for truly solidifying the information, the follow-up is apply, apply, apply. Application of knowledge gained should always be the end goal in learning whatever you do, and that's exactly what Phase 7 of the Norudit Sequence, the Capstone, is built to do. This also combats a major issue students experience in learning: the "illusion of knowing." Basically, it's believing you know something you don't because it looks or feels familiar, with the major culprits being rereading and highlighting. Learn more about the illusion of knowing.

The best move

The best rewards often tend to come from the narrower path no one wants to walk. Hard work never guarantees success, but hard work along with smart work, more often than not, yields a significantly better result than seeking the easy way out.

Interest-based mapping: the concept, wrapped in what you already love

The core idea is simple: it's all about relatability. You learn better when it's related to something you already know and enjoy, because it's easier to focus, and easier to understand, since you can relate what you're learning to a concept you already grasped.

You've probably heard, or said, something along these lines: "Why do I need to learn this? I'll never use it in real life." The problem is usually that educators focus solely on the material and miss a crucial step in teaching: relating it to something the student is interested in. This is the basis of Phase 2 of the Norudit Sequence. In the settings page, students can list up to 15 interests, and the AI is aware of them but only uses one where it genuinely fits, instead of forcing it. Where none fit, it falls back to a general analogy most people would relate to.

For example, if a student likes to read manhwa (Korean comics) and they're being taught about taxes in economics, a lesson could ask the student how they'd handle taxes if they were transmigrated into a medieval kingdom and expected to run its finances.

This is intrinsic motivation, not the engagement farming (streaks, XP, badges) I mentioned at the start of this post. Engagement farming manufactures a feeling of progress that only exists inside the app. Interest-based mapping doesn't manufacture anything, it simply engages with an interest that already exists and links it to the concept being taught. I go deeper into why that difference actually matters for outcomes in Why Grades Shouldn't Be the End Goal of Studying.

Grounded in your material: the principle that makes the other four trustworthy

Onto the last thing I'll address in this post, a problem that's become more prevalent in recent years with the rise of AI, or more specifically Large Language Models (LLMs) like ChatGPT and Claude. This problem is specific to AI-based study methods. AI makes mistakes. AI, best explained right now, is like a toddler with the intelligence of a genius and the knowledge of an encyclopedia. Although that's true, AI hallucinates due to malformed or inaccurate data in its training set, or from inaccurate web sources.

The answers from AI should never be fully trusted. So does that mean never use AI? Quite the opposite. I'd even go so far as to say AI is an invention on the same level of importance as the computer and fire. The solution isn't to push AI aside, but to find ways to mitigate the issues with current models.

I'm not going to get into all the areas that need improving, but for learning specifically, the idea is making sure the AI is grounded in reliable sources, like your textbook, and that it uses ONLY that source for its responses, rather than relying on open web search or its training data, except where research specifically calls for it. If you'd like to learn more about Norudit's attempt at tackling this issue, read more here.

If you'd like to give these solutions a go: Start your first class →.

Read more