Why spaced repetition actually works — and how to use Lemnly to ride it

Almost every language learner has had the same uncanny experience: a word you knew last week, gone today. You looked it up, you used it in a sentence, you swore you would remember — and then the brain quietly recycled it. It is tempting to take this personally. It is not personal. It is the forgetting curve, and it is the single best lever you have for learning.

The forgetting curve, briefly

Hermann Ebbinghaus, in 1885, sat down and memorised long lists of nonsense syllables, then tested himself at intervals to plot how fast he forgot them. The shape he found is approximately exponential: most of the loss happens within the first 24 hours, then the curve flattens. The good news buried in this discouraging finding is that each successful retrieval flattens the next curve. Forgetting is not the enemy. Forgetting and then recalling is what teaches you.

Memory is not a recording. It is a muscle, and the way to train it is to make it strain just before it would have failed.

Why interval guessing fails

Most flashcard apps — including some that should know better — schedule cards on a fixed-interval ladder: 1 day, 3 days, 7 days, 14 days. This ignores the fact that cards are not interchangeable. The Italian word cane ("dog") and struggente ("poignant") behave nothing alike in your head. Fixed intervals over-rehearse the first and under-rehearse the second. You waste time on cards you’d know in your sleep, and you lose cards you almost had.

Worse, fixed intervals assume your forgetting curve is the same as the next person’s. It isn’t. Sleep, age, mood, time of day, how emotionally loaded the word is — all of these change how long a card sticks. A scheduler that ignores per-card and per-user signal is playing the average game on data that isn’t average.

What FSRS does differently

FSRS (Free Spaced Repetition Scheduler) does three things that older algorithms like SM-2 do not:

• It models per-card difficulty. Two parameters — stability (how long the memory lasts) and difficulty (how hard the card is for you) — are updated after every review. The model learns which of your cards are sticky and which are slippery.
• It targets a specific retention probability. Typically 90%. That number is a knob you can turn. Higher retention means more reviews; lower means fewer. You choose your tradeoff instead of inheriting someone else’s.
• It is trained on real data. Not a clever guess from 1987, but a fit to billions of review logs across the open-source SRS community.

The practical effect, in numbers

In our own beta data — admittedly small — we saw the same retention as SM-2 with roughly 28% fewer reviews. That is the difference between five minutes a day and seven. Over a year, that is enough to read an extra novel — or to give yourself two extra weekends off without bankrupting your deck.

The pattern shows up across every published comparison: FSRS hits a target retention with fewer reviews, especially as your card count gets larger. By the time you’re past 5,000 active cards, the difference is huge — about a third less time per session for the same retention.

How to set your retention target in Lemnly

Lemnly lets you choose between 80% and 98% target retention, per-deck. The right number depends on what you’re using the card for.

• 80% — minimum viable recall. Right for passive recognition vocabulary you mostly meet in reading. Fewer reviews, you’ll still recognise the word when you see it.
• 90% — the default. Right for most learners, most of the time. Good balance between effort and confidence.
• 95% — high stakes. Right for upcoming exams, professional vocabulary, or words you actively use in conversation.
• 98% — overkill, almost always. Right for the dozen highest-frequency function words and almost nothing else.

A common pattern: 90% on your main "reading vocabulary" deck, 95% on a small "speaking vocabulary" deck, 80% on a "nice to recognise" deck for rare words from books you read once.

The four-button question

After every card, Lemnly asks: Again, Hard, Good, Easy. Most learners use this badly. Here’s how to use it well.

• Again — you genuinely didn’t recall it. Not "I had to think for a second." Not "I almost had it." Press Again only when you’d have been unable to use the word in a sentence cold.
• Hard — you recalled it, but it cost you. You searched, you frowned, you guessed and got lucky. The word is on the edge of your recall and the schedule should not stretch much.
• Good — you recalled it without strain. This is most cards on a good day. Don’t feel obliged to upgrade to "Easy" just because you’ve seen the word a few times.
• Easy — you recalled it instantly, no effort. Use this sparingly. Pressing "Easy" tells the scheduler to skip the card for a long time, and you’ll miss the chance to keep stability climbing.

Calibration matters more than honesty. If your "Good" really means "Hard," the scheduler will over-stretch and your retention will crater. Treat the buttons as a forced-choice signal to the algorithm, not a self-assessment.

Where most apps still get it wrong

Even with FSRS underneath, an app can sabotage you in three ways:

1. Card glut. Adding every word you ever met means you will drown in reviews and quit. Lemnly only proposes the words you actually do not know, by checking against your own vocabulary and a shared cache. The hard work is in the curation, not the collection.
2. Brittle cards. A bare word on one side and a single translation on the other is a context-free coin toss. We attach the original sentence and a part of speech so the recall is anchored. A word memorised in context is roughly twice as durable.
3. Punishing the user for living a life. Streaks that break on a missed day train people to lie. Lemnly’s streak counts the weeks you reviewed at least four days — generous, but real. Skip Sundays. Skip your wedding. The streak survives.

A 30-day plan to actually feel the difference

Try this for one month. If you don’t feel a difference, you can have your money back — both of the pounds the free plan cost you.

1. Days 1–2: Set up. Pick one target language. Set 90% retention. Import one source — a short article via URL is the easiest first move.
2. Days 3–14: Review every day. Five minutes. Calibrate the buttons. Don’t add new cards.
3. Day 15: Import your second source — something longer this time, like a chapter of an EPUB. Review the preview, trim half the proposed cards.
4. Days 16–30: Daily review. Once a week, add 10–20 new cards from whatever you’re reading.

At day 30, look at your stats. If you’ve done the daily five minutes, your active deck will have 60–80 mature cards with 90%+ retention. That’s 60–80 words you genuinely own. Two years of that pace is the difference between "I studied French" and "I read in French."

The takeaway

Spaced repetition works. It worked for Ebbinghaus and it works for the sleep-deprived parent reviewing Italian on the U-Bahn. The job of a good app is to put a modern scheduler in front of you without making you study its UI. The job of the learner is to show up for five minutes a day and be honest with the buttons. That is the whole product, really. The rest is reading.