memori
The two-question test
Before a card enters your deck, ask two things. Could a stranger answer this correctly with only the front? And is there exactly one answer I would accept? If either answer is no, the card will fail repeatedly and you will blame your memory. Nearly every rule below is a specific way of passing this test.
The reason this matters more than it sounds is compounding. A card is not a one-off cost. Add it to a spaced repetition deck and you have committed to answering it perhaps eight or ten times over the next two years. A bad card doesn't waste thirty seconds; it wastes thirty seconds repeatedly, forever, while also generating the feeling that the method isn't working.
Most of what follows descends from Piotr Woźniak's Twenty rules of formulating knowledge, written in 1999 and still the best thing on the subject. What is added here is the rewriting — seeing the broken card next to the fixed one is what makes the rules transferable to your own deck.
The nine rules
1. One card, one fact
The minimum information principle: each card should test the smallest unit that is still meaningful. Compound cards fail as a unit — forget one clause of six and you mark the whole card wrong, which drags six pieces of knowledge back to a short interval because of one.
Mitochondria
Double-membraned organelle; site of oxidative phosphorylation; has its own circular DNA; inherited maternally; inner membrane folded into cristae.
Which organelle is the site of oxidative phosphorylation?
The mitochondrion.
Mitochondrial DNA is inherited from which parent?
The mother.
What are the folds of the inner mitochondrial membrane called?
Cristae.
The heuristic: if your answer contains the word "and", or a semicolon, or a bulleted list, you are looking at more than one card.
2. The front must be a question with one correct answer
A bare term on the front is ambiguous. Shown the word "entropy", are you meant to produce a definition, a formula, its units, or an example? You'll produce whichever comes easiest, mark yourself correct, and never notice that you can't produce the others.
What is the function of the liver?
Detoxification, bile production, glycogen storage, plasma protein synthesis, urea cycle…
Which organ converts ammonia into urea?
The liver, via the urea cycle.
Which plasma protein, synthesised by the liver, is the main determinant of oncotic pressure?
Albumin.
A useful test: imagine marking your own answer. If you'd have to decide whether what you said was "close enough", the question is too open.
3. Scope the question
Your deck eventually contains material from several subjects, and a question that was obvious in context becomes ambiguous in a mixed queue at 11pm.
What is a stack?
A last-in-first-out data structure.
Data structures — which structure has last-in-first-out access order?
A stack. (Contrast: a queue is first-in-first-out.)
A two- or three-word prefix costs nothing and removes an entire class of failure. Note also the parenthetical contrast on the back — it costs you nothing to review and pre-empts the confusion in rule 5.
4. Understand it before you schedule it
This is the rule with the worst effort-to-compliance ratio, because breaking it is invisible for about three weeks and then extremely expensive.
A card you don't understand is not a memory problem, it's a comprehension problem wearing a memory problem's clothes. You will fail it, re-fail it, watch the interval collapse to a day, and slowly come to dread your reviews. Memorising the formula for a t-test without understanding what it compares means you'll produce the formula and still be unable to use it.
The fix is upstream: understand it first, from a lecture, a worked example, a textbook, or by talking it through with someone until it clicks. Only then write the card. Active recall preserves understanding; it doesn't create it.
5. Make interfering cards distinguishable on the front
Similar cards actively fight each other. This is interference, and it is the single most common reason a specific pair of cards refuses to stick — you know both answers, you simply can't reliably attach the right one to the right question.
Which cranial nerve supplies the lateral rectus?
CN VI (abducens).
Which cranial nerve supplies the superior oblique?
CN IV (trochlear).
Which cranial nerve supplies the lateral rectus — the muscle that abducts the eye?
CN VI, the abducens. ("LR6")
Which cranial nerve supplies the superior oblique — the muscle that turns the eye down and in?
CN IV, the trochlear. ("SO4")
Two changes did the work. Each front now carries the distinguishing feature (what the muscle actually does), and each back attaches the answer to something memorable — "abducts / abducens" is a genuine etymological hook, not an arbitrary mnemonic. When you can't find a hook, an explicit contrast on the back ("not to be confused with…") is the next best thing.
6. Don't memorise lists as lists
Enumerations are the worst thing you can put on a flashcard. "Name the seven characteristics of living organisms" requires all seven, in a deck that penalises you for missing one, and the recall is serial — lose item three and items four onward often go with it.
List the seven characteristics of living organisms.
Movement, respiration, sensitivity, growth, reproduction, excretion, nutrition.
Biology — what does the mnemonic MRS GREN stand for? (Give the first letters only.)
Movement, Respiration, Sensitivity, Growth, Reproduction, Excretion, Nutrition.
Which characteristic of living things does sweating in response to heat demonstrate?
Excretion — and sensitivity, since it's a response to a stimulus.
Two moves: give the list a single retrievable handle (a mnemonic, an acronym, a memorable ordering) and then build recognition cards that ask you to apply individual members. You end up knowing the list and being able to use it, which the original card never tested.
7. Write "why" and "when" cards, not just "what"
Most decks are almost entirely nouns. Most exams — and all real use of knowledge — ask verbs. This is the cheapest available upgrade to a deck's usefulness.
What is the average-case lookup complexity of a hash table?
O(1).
Under what conditions does hash table lookup degrade to O(n), and why?
When many keys hash to the same bucket — a poor hash function, a pathological key distribution, or adversarial input. Lookup then has to scan that bucket's chain linearly.
Keep the "what" card too — you do need the O(1) fact. But a deck with only "what" cards produces someone who can recite the complexity table and still can't tell you why their program got slow.
8. Use your own words and your own examples
Copying a textbook sentence onto a card is transcription, and transcription is the least effective thing you can do with material. The generation effect — that information you produce yourself is remembered better than information you read — has been reported since Slamecka and Graf's 1978 experiments.
Practically: rephrase the answer as you'd explain it to a friend, and put a concrete example on the back, ideally one from your own life or coursework. "Confirmation bias" with a definition is a card. "Confirmation bias" with the definition plus the time you only read reviews of the laptop you'd already decided to buy is a card you'll still have in five years.
9. Date it and source it when the fact might move
Some facts have a shelf life: clinical guidelines, tax thresholds, API behaviour, "the current best-performing model". Put a short source and date on the back. Two years later you'll be able to tell at a glance whether the card is still true, instead of confidently reproducing a superseded guideline in an exam.
This costs a few seconds and applies to maybe one card in twenty. Skip it for the periodic table.
Writing cards in your own words is the right approach and also the reason most people stop making cards. One workable compromise: understand the material through conversation first, then let the phrasing be drafted for you and edit it. In Memori, tapping Save Memori on an explanation you've just worked through produces draft cards from that conversation — which are in the vocabulary you were just using, and which you preview and edit before they're saved.
When to use cloze deletion
A cloze card hides part of a sentence and asks you to supply it: "The powerhouse of the cell is the […]". Use cloze when the surrounding sentence is genuinely load-bearing context — sequences, formulae, syntax, quotations. Avoid it when the sentence gives the answer away, which is the standard failure.
Cloze is popular because it is fast — you can turn a sentence into a card without rephrasing anything. That speed is also its danger: it makes it effortless to produce cards that test nothing.
Good uses: steps in an ordered process, chemical or mathematical expressions where structure matters, language sentences where grammar supplies the cue, quotations and legal wording where the exact phrasing is the point.
Bad uses: any sentence where the remaining words uniquely determine the missing one. "The 1919 Treaty of […] formally ended the First World War" is not a memory test; it's a crossword clue with one possible answer. If you can solve it without knowing the material, so can the version of you who has forgotten it.
A quick check: read the cloze with the blank and ask whether someone who had never studied this could guess it. If yes, either delete more of the sentence or turn it into a proper question.
Fixing cards that never stick
A card you have failed six or eight times is called a leech, and it is a signal about the card rather than about you. The correct response is to rewrite or delete it, not to review it harder. A handful of leeches can consume a disproportionate share of your total review time while teaching you nothing.
Diagnose before you rewrite. Run through these in order:
- Does it ask for more than one thing? Split it. This fixes maybe half of all leeches on its own.
- Is there a sibling card it keeps swapping with? If you can produce both answers but attach them to the wrong questions, it's interference. Rewrite both fronts to carry the distinguishing feature, as in rule 5.
- Do you actually understand it? Try explaining it aloud. If you can't, the card was never the problem. Go and learn the thing, then rebuild.
- Is it arbitrary? Some facts have no hooks — a random date, a code, a name. These need a deliberate mnemonic or an image. Arbitrary information is legitimately harder, and pretending otherwise just produces guilt.
- Do you actually need it? Genuinely ask. A surprising number of leeches are cards made out of completeness rather than need. Deleting one is not failure; it's curation.
How many cards to make
Fewer than you think. A dense textbook chapter typically yields fifteen to thirty cards worth keeping. If you're producing a hundred, you're transcribing the chapter rather than selecting from it — and since each card commits you to roughly eight to ten reviews over the following years, over-carding is how decks become unsustainable.
The selection question that works: would I be annoyed with myself for having forgotten this in six months? Not "is this true", not "was this on the slide" — would forgetting it actually cost you something. Most sentences in most chapters fail that test.
It also helps to make cards after a retrieval attempt rather than during reading. Read the chapter, close it, write down what you remember, and only then build cards — for the gaps. Anything you recalled unaided doesn't need a card yet, and this single change typically cuts card production by half while improving what remains.
A ten-minute deck audit
If you already have a deck that isn't working, this is the fastest diagnostic:
- Sort by lapse count and look at your worst 10 cards. Not to review them — to read them. You will almost always find the same two or three patterns repeated across your whole deck.
- Count the words on the back of those cards. Anything over about 20 words is a compound card pretending to be one card.
- Look for near-duplicate fronts. These are your interference pairs. Fix them together, never separately.
- Read ten random cards and try to answer them cold. Any card whose front leaves you thinking "answer what, exactly?" fails rule 2.
- Delete ruthlessly. Aim to remove 10% of the deck. The reviews you free up buy you back real time every day for years.
An audit usually surfaces the same fix repeated across dozens of cards — add a subject prefix, shorten every answer, add an example sentence to each. Doing that by hand is exactly the kind of chore that stops people auditing at all. Memori's Magic Reformat lets you select a batch of cards and apply one instruction to all of them, with before-and-after previews so you can see what changed before committing.
Where generated cards fit
AI is genuinely good at the mechanical half of card writing — splitting material into single facts, phrasing unambiguous questions, keeping answers short. It cannot know what you already find easy, and it will occasionally state something false with total confidence. Treat generated cards as drafts that require reading, not as output.
The realistic division of labour:
| Task | Delegate? | Why |
|---|---|---|
| Splitting a paragraph into atomic facts | Yes | Mechanical, rule-driven, tedious |
| Rephrasing a statement as a question | Yes | Same |
| Formatting a batch consistently | Yes | Pure chore, no learning content |
| Deciding what's worth remembering | No | Depends on your course, exam and existing knowledge |
| Checking the fact is correct | No | Confident errors are the main failure mode |
| Doing the retrieval | Never | That's the part that constitutes learning |
A practical rule: if a generated card enters your deck without you having read it, you've automated the wrong half. The fuller argument is in how to use AI to study without cheating yourself.
Frequently asked questions
How many flashcards should I make per chapter?
Fifteen to thirty for a dense chapter is a normal, sustainable range. If you're at a hundred you're transcribing. The best way to control this is to make cards after a closed-book recall attempt, so you only build cards for the things you actually failed to remember.
Should the front be a question or just a term?
A question. A bare term doesn't tell you what you're being asked for, so you'll answer with whatever surfaces first and mark yourself correct. Questions make the retrieval target explicit, which is the whole point of the exercise.
Why do I keep forgetting the same card?
Because it asks for too much, it's interfering with a similar card, or you never understood the underlying idea. All three are fixed by rewriting, not by reviewing more. Reviewing a broken card more often just spends more time confirming it's broken.
Should I make reverse cards?
Only where you genuinely need both directions. For vocabulary you intend to produce — speaking, writing — yes, both directions. For recognition-only vocabulary, or for facts like "who discovered X", the reverse card is usually meaningless and doubles your review load for nothing.
Are images worth adding?
For anything spatial or visual, very much so — anatomy, chemical structures, graphs, maps, circuit diagrams. A labelled diagram with one label hidden is often the single best possible card for that material. For abstract propositions, a decorative image adds review time without adding a cue.
How long should a card take to answer?
A few seconds. If a card routinely takes you thirty seconds of reconstruction, it's a mini-essay and should be split. The exception is deliberate synthesis cards — "explain why X causes Y" — which you might keep a small number of, knowing they cost more.
Where this comes from
- Woźniak, P. (1999). Effective learning: Twenty rules of formulating knowledge. SuperMemo.
- Nielsen, M. (2018). Augmenting Long-term Memory. An essay on the practice of writing cards for spaced repetition.
- Slamecka, N. J., & Graf, P. (1978). The generation effect: Delineation of a phenomenon. Journal of Experimental Psychology: Human Learning and Memory, 4(6), 592–604.
- Karpicke, J. D., & Roediger, H. L. (2008). The critical importance of retrieval for learning. Science, 319(5865), 966–968.
- Anderson, M. C., Bjork, R. A., & Bjork, E. L. (1994). Remembering can cause forgetting: Retrieval dynamics in long-term memory. Journal of Experimental Psychology: Learning, Memory, and Cognition, 20(5), 1063–1087. (On interference between similar items.)
- The Anki manual's sections on leeches and card templates, for the operational definitions used here.
