12 min read

The Deck That Rewrites Itself

noruditlearning sciencespaced repetitionmemory

Spaced repetition is the most proven idea in all of studying, and the most quietly broken in practice. Two things kill it, and a fixed deck of flashcards walks straight into both. Here is how Norudit's review deck is built to survive them.

Start with what's right

Spaced repetition works. Schedule each thing you're learning to come back just as you're about to forget it, and a small daily effort holds thousands of facts for years. That is the spacing effect Ebbinghaus (1885) first charted and Cepeda et al. (2006) confirmed. This isn't a theory; it's how much of the world's medical workforce got through their exams. Anki, the tool most people mean when they say "flashcards," deserves real credit for proving it at scale.

Norudit runs on the same foundation. A modern scheduler decides when each concept comes back: FSRS, built on the memory model of Ye et al. (2022), which can be fitted to your own forgetting curve from your review history. A mastery model decides what comes back first, putting your weakest concepts at the front.

So far, so standard. The problem is what happens to a flashcard over time.

Failure one: the card stops testing you

Here is the trap nobody warns you about. The card never changes. Show yourself the exact same card fifty times and, somewhere along the way, something shifts: you stop retrieving the idea and start recognising the card, its shape, its wording, where the answer sits, the little coffee-stain of familiarity it carries. You flip it, you feel the click of recognition, you mark it "good." The schedule stays green. And underneath, the knowledge is hollowing out, because recognising a card you've seen fifty times is not the same act as knowing the thing it points to. You have memorised the flashcard. You have not kept the concept.

This is why so many people grind a deck for months and then freeze in the exam: the deck was testing card-recognition, and the exam tests knowledge, and those turned out to be different skills.

The fix: concept-varied spaced repetition

In Norudit, every time you answer a card, the card is thrown away and a new one is built for that same underlying concept. The concept is fixed; the surface is regenerated. This is concept-varied spaced repetition: the same proven spacing schedule from the section above, paired with a card that never asks the same way twice. And the kind of new test depends on what kind of thing you're learning, because concepts and facts fail in different ways:

Concretely, every card type regenerates:

The next time a concept is due, you meet it as a stranger. There is no surface left to pattern-match against, so the only thing in the room that can answer the card is the concept itself. This is not a gimmick. Varying the conditions of retrieval is one of the best-established findings in the science of memory, one of Robert Bjork's "desirable difficulties" (Bjork 1994; Bjork and Bjork 2011) and a direct descendant of the retrieval-practice and interleaving research of Roediger and Karpicke (2006) and Rohrer and Taylor (2007). It's just that fixed-card tools have no way to do it, and a system that regenerates every card does it for free.

The principle

A fixed card eventually tests whether you recognise the card. A card that rebuilds itself always tests whether you know the concept.

It changes how hard, not just how it looks

The surface is not the only thing that adapts. How demanding the rebuilt card is tracks how well you actually know the concept, read from two signals at once: how you just graded this card, and your accumulated mastery of the concept underneath it.

So a concept you are still finding your feet on is treated gently, and one you have genuinely mastered is stretched, without you having to decide any of it. A fixed deck cannot do this. A system that rebuilds every card and tracks your mastery per concept does it automatically, every time the card comes back.

Failure two: the backlog becomes a wall

The second way spaced repetition dies has nothing to do with cards and everything to do with being human. You miss a week (you're ill, it's exam season, life happens) and you open the app to 347 cards due.

That number is not a to-do list. It's a wall. Most people take one look, feel the weight of it, and close the app. Then they miss another week, and the wall grows, and the guilt grows with it. More Anki habits die in this shame spiral than ever died of boredom. The schedule was perfect; the human bounced off it.

So Norudit's review session never shows you the wall. When a backlog exists, review is sequenced instead of dumped:

That design is doing three quiet things at once:

  1. The first sitting is always finishable. Opening the app after a bad week feels like taking a step, not receiving a sentence. Thirty is a number you can look at without flinching.
  2. A truncated session is still the optimal session. Because the order is weakest-first, even if you only ever do those first 30 cards, they were the 30 that mattered most. You are never punished for doing less than everything.
  3. One neglected class can't hold the others hostage. The cap is per class, so a backlog in chemistry burns down over a few bounded days while your maths review stays perfectly current.

Forgiveness is engineered in, because a spacing schedule only works if you keep showing up, and you only keep showing up if coming back is never made to hurt.

The best review schedule in the world is worthless if the wall it builds makes you quit. So we don't build the wall.

Calibrated to your forgetting curve, after about 400 reviews

There's one more layer, and it's the quietest of the three. Everyone forgets at a different rate: the research parameters FSRS-6 ships with are a very good average human, but you are not an average human, and neither is anyone else.

So the scheduler personalises in two stages. From day one, every single review updates each card's memory model from your grades. Struggle with a card and its intervals compress, breeze through it and they stretch. And once you've logged about 400 reviews, you can go one step further: a single tap runs the full FSRS optimiser on your own review history and refits all of the algorithm's parameters to your measured forgetting curve, scored against your actual recall record, not a population average. From that moment, every interval in every deck is computed with weights trained on you. (Under 400 reviews the maths doesn't have enough signal to fit honestly, so the button politely waits; the research defaults are the safe start, not a compromise.)

Rebuilt cards decide what you retrieve; the fitted curve decides when. Both halves personal, along with everything else the academy remembers about you.

Why the deck stays locked until you've earned it

Here's a choice that surprises people: for a new topic, the review deck does not exist yet. There are no cards to grind and nothing to open, not until you have passed the Feynman gate (Phase 3) and finished your first free recall (Phase 4). Only then is the deck built, generated from that topic's material in your own uploaded documents, now that you've shown on a blank page that you both understand it and can pull it back. That lock is deliberate, and it comes down to two plain reasons.

Reviewing something you never understood just drills the wrong thing. Spaced repetition is a maintenance machine: it faithfully keeps whatever you feed it. Feed it a concept you never actually grasped and it doesn't close the gap; it sets the gap in concrete, resurfacing a hollow or half-wrong idea again and again until the mistake starts to feel like fact. So understanding has to come first, and that is exactly what the Feynman gate checks: if you can't explain it simply, you don't yet know it. There is no point maintaining a model that was never right in the first place; the gate is the only honest way to check the understanding is really there before review begins.

And there's nothing to keep if the memory never formed. Passing the gate proves you understand the idea today. It does not prove the memory took hold. That is what the blank page of free recall settles: pull the whole topic back out of nothing, and what you can reconstruct is what actually stuck. A review schedule pointed at a memory that never formed is just maintenance on an empty room. So the deck for a topic isn't generated until that topic has been understood and recalled at least once. It's still built from your own material, but only after the blank page has proved there's a real memory there worth maintaining.

Put those two together and the lock does something kind as well as strict: your deck only ever holds concepts you have already understood and already recalled at least once. You never open it to a pile of cards for things you haven't properly met yet. Every card is something you have a fair shot at. That is a second layer of the same mercy as the sequenced backlog above: that one saves you from too many cards; this one saves you from unearned ones. Review should feel like keeping something you already own, never like cramming something you don't.

How it connects to everything else

The deck doesn't appear from nowhere, and it doesn't run in isolation. A topic's review deck is built the moment you finish its first free recall (the blank-page test you take right after passing the Feynman gate), generated from that topic's material once you've earned it, so you never review a topic you haven't first understood and recalled. And the weakest-first ordering is fed by the same mastery ledger that the exam diagnoses into and that the memory layer maintains. It is one turn of the larger learning flywheel, the part that keeps what the rest of the sequence built.

Every tool that feeds it is in the full tour of the academy. Start your first class →

Sources

Read more