You answer from memory. An editor marks it.
Two of the most robust, best-replicated findings in the learning sciences: effortfully producing an answer from memory builds far more durable retention than re-reading it, and spacing those retrievals out beats massing them. Everything in LearnReally is those two facts, engineered — plus one honest correction: your own feeling of “I know this” is a bad judge, so a grader checks.
Every card starts as an empty field.
A due card shows you a prompt and nothing else. You type — or dictate — what you remember. No multiple choice, no flipping to peek first. Recognizing an answer you're shown is the weakest form of memory; producing one is the strongest, and the struggle itself is what does the strengthening. A card that felt effortful and came out right is a card that just got more durable.
Graded against the reference — never against the model's imagination.
When your agent or the Builder writes a card, it stores a reference answer and the key points that matter. The grader judges your recall against that stored reference only — it is explicitly forbidden from “correcting” you with its own knowledge. That's the deliberate defense against a model confidently marking you wrong with something it invented.
The grade reads like an editor's note in the margin: a score, what you nailed, what you missed and what the reference actually says. Meaning counts, wording doesn't — a paraphrase that captures the idea is a recall; a fluent sentence that misses it is not. If you used a hint, the suggested rating caps at “Hard” — assisted recall is real, but it isn't the same evidence. The final call on every rating stays yours.
Why not just self-rate? Because self-judgment tracks how fluent something feels, not whether you'll recall it next week — that's the fluency illusion, and it's one of the best-documented failure modes in metacognition research. Self-review is still a full session type (free on every plan, and it never spends an AI credit); the graded Recall loop exists for the moments you want the honest answer.
FSRS decides when — you just show up.
Scheduling runs on FSRS, the modern open spaced-repetition algorithm (the same family modern Anki uses). Every review updates a per-card memory model, and the next review lands right before the model predicts you'd forget. Intervals grow when you recall and reset toward relearning when you don't. We don't hand-roll interval ladders and we don't sell a proprietary “algorithm” — the science here is a commodity, honestly applied.
New material arrives in course order and in bounded daily batches, so a 400-card course doesn't bury you on day one. And when nothing is due, the app says so and stops you. No streaks, no daily goals: they measure whether you showed up, not whether you remember. Memory is supposed to be quiet.
When you do want an extra pass — the night before an exam, say — every deck has Practice: the same session over the whole deck, graded as usual, but off the record. Practice never moves a card's schedule and never counts against the daily cap, so cramming can't corrupt the spacing model that makes the schedule work.
Self-rated vs. proven.
Self-review sessions — reveal and self-rate, no typing — are first-class: they're real reviews and FSRS uses them. But the record stays honest about what kind of evidence each review is. A card counts as proven only when a produced answer — typed or spoken — has witnessed it. Your stats carry that number, and nothing nags you about the difference: the number itself is the motivation.
After the verdict, the conversation.
The grader never has the last word: it suggests a rating, and the four buttons under the answer are yours — what you press is what the schedule follows. If the feedback raises a question, ask it right there: the deep-dive defines terms, adds background, and explains the idea from other angles — anchored to the card, never contradicting it — up to a few turns per card. And anything worth keeping — a fragment you keep missing, a tutor answer that finally made it click — becomes its own card: click the line, and the card opens as a preview before it is saved, scheduled from then on like any other.
The claims you won't find here.
No learning-styles matching (debunked). No decorative images or gamified dopamine — decoration measurably lowers learning, and streaks shift attention from the material to the self. No “learn anything 10× faster” promises: cards build durable memory of exactly what they test, and far transfer mostly doesn't happen — anyone promising otherwise is selling something. The whole loop rests only on high-confidence findings; you don't have to trust any fragile study to trust the product.
Want the session modes in practice — Recall vs. Self-review, what counts, session choices? See Getting started and the FAQ.