It all starts with your own dictionary
KYW’s main job is to help you grow your vocabulary steadily without losing what you already learned. That is why the words are yours: you add the ones you don’t remember yet but want to know, and the app works only with them. A card consists of the word itself, its definition, an optional transcription, tags, synonyms and usage examples. The definition is what you see during a session; the word is what you have to type.
An account can hold as many languages as you like, named however you like: “German”, “Japanese”, “Latin”, “anatomy terms”. The dictionaries are fully separate – own words, own tags, own statistics, own schedule, nothing gets mixed. Switching between them takes one click.
You don’t have to type the dictionary by hand. The “Add” tab has a button that copies a ready-made prompt: give it to any chatbot together with your word list or topic, and paste the answer back. It is the fastest way to get fifty cards with sensible definitions and examples.
The second way is a pack. It is a named list of words somebody has already put together: a teacher for a class, a colleague for a team, or a stranger who published it to the shared library. A pack often carries the text those words occur in – then a word arrives out of a phrase rather than out of nowhere. The words you take are copied to you for good: you can edit them, and the author can neither take them back nor change them on your side.
How a single card goes
You see the definition. You type the word and press Enter. There is no peeking before that – this is the difference from cards you flip, where the answer is shown before the check and rating your own knowledge is left to your conscience. Here the word is either typed or it is not. As a result you learn how the word is spelled, not how it looks, and the second is worth more.
A typo is not treated as ignorance. Swapped letters or a missing double consonant differ from typing another word altogether, and the app accounts for that difference. A rating is suggested automatically, but the choice stays yours: four buttons from “didn’t know” to “easy”. A rating given right after typing the word is more accurate than one given after simply looking at the answer.
Review: rescue what is about to be forgotten
What is due comes first, and the order there is not random: the words are sorted by recall probability, so whatever has sagged most is shown earlier. The stage runs in two passes, and the first is the general one: every overdue word, once each.
Words answered on the first attempt go into the schedule and leave the session – they stop taking your time. The missed ones are collected separately and go into a second pass, this time shuffled, so the sequence is not what you memorise. That is the sifting: effort concentrates where it is needed. A word failed in both passes is not dropped from the session – it is finished off in the third stage.
New words: learn first, schedule after
First the introduction: the word is shown together with its definition, with nothing to type. The goal is to see the link once, not to memorise it at a glance.
Then comes a general pass over the whole set. If it goes through without a single miss, the set moves into the schedule. If not, the missed words spin in rounds until none are left, and then the general pass starts over from scratch. The point is not the number of repetitions but the threshold: a schedule only makes sense for a word you reproduce on the first attempt in the general flow, not right after you have seen it.
Finishing off and getting ahead
The third stage starts with your failures – the words you named in neither pass of the review. They are finished off here, while the session is still running and the context has not cooled. Then the early ones are mixed in: not formally due, but holding weaker than the rest, and the first shown are those whose recall probability has already sagged – regardless of the calendar date.
Only that second half is optional – and it is the half usually skipped. The gain is simple: a word caught at 78% probability costs one attempt. The same word left until it fails will need drilling from scratch and will reset the interval you had accumulated. Five minutes here are cheaper than starting over a week later.
Inside a stage the order is not random either
Every card inside a session has its own step – how many other cards until the word is shown again. A miss shrinks the step and the word comes back quickly; a confident answer stretches it and the word recedes. Readiness is what has passed since the last showing divided by the step: a word with a bigger step ripens longer and is drawn less often. The next card is taken by lottery weighted by readiness squared: usually the ripest one, sometimes its neighbour, so the order never becomes something you learn by heart.
While many unfinished words are in play, the pressure stays high and new ones barely come in. As the set consolidates, the pressure falls and the session opens up to fresh words again.
In how many days the word returns
To determine that, KYW uses FSRS-6, a modern memory model that computes spaced repetition. The algorithm aims for the moment just before forgetting: repeating too early wastes time, too late returns the word to a state where it has to be learned again.
For every word the model keeps two quantities: how firmly it is held and how hard it is for you specifically. From those it derives the probability of recalling the word today. The word counts as due when that probability drops to the retention level you set – 90% by default. Hence the growing intervals: the firmer the word, the longer the pause.
Drilling is handled separately. In the classic scheme every answer is an event for the model, and ten mistakes inside one session would bury the word for months. In KYW a session produces one record per word: in-session misses are folded together with a decaying weight, each next mistake counting half as much as the previous one. A word that came hard starts more cautiously, but not from zero.
You can see what is happening
Statistics answer the questions that appear after a month: how many words actually hold rather than merely exist; is accuracy growing; is a pile-up building for the day after tomorrow. The load forecast shows how many reviews are coming in the next few days – it tells you whether to take new words today.
Stubborn words are collected separately – the ones that keep failing. Usually that means the definition is poor or the word was learned without context; such a card is easier to rewrite than to grind. You can share your stats with a link – the page opens without signing in and shows nothing but numbers.