product · spaced recall

A deck that writes itself. From your own errors.

Every wrong attempt becomes a card. Every low-scored explain-back becomes a card. Cards return spaced by SM-2 — the tricky ones on the day you're about to forget them.

SM-2, since 19909s median review · offline-capable

Median review time

Confidence tag → flip → four-button grade.

auto

Card generation

Every wrong attempt and every explain-back < 4.

SM-2

Scheduling algorithm

1990. Battle-tested. Every good SRS descends from it.

the reason this exists

Retention halves in a day. Then again. Then again.

Drag the slider — day 14, no review: 15% retention. Day 14, five reviews on the SM-2 schedule: 92%. Same brain. Different schedule.

Drag time — day 3 since you first learned it

Real math. R(t) = e^−t/S. Each review multiplies S by ~2.2 — the same relationship every spaced-repetition system in the world is built on.

the SM-2 schedule

Each successful review stretches the next interval.

A card you nailed comes back later — a card you fumbled comes back tomorrow. The math tunes itself against your grades so the queue always presents you with cards on the day you’re about to forget them.

a card you nailed comes back later — a card you fumbled comes back tomorrow

the reviewer

Twenty cards, two keystrokes each.

A real capture of the screen you spend the minutes in. Space turns the card; 1 to 4 grade it. You never have to reach for the mouse, which is the difference between a review habit that holds and one that quietly stops.

The Brainback recall reviewer: a segmented queue bar, a card showing a prompt, and the keyboard hint to turn it over. — brainback.chat/app/recall · captured from production

the algorithm, in the open

The scheduling code, verbatim.

This is the entire scheduling function. Twenty-eight lines. Deterministic. Testable. If a card’s next-due date ever surprises you, this function is the answer.

lib/spaced-recall/schedule.ts

typescript

// lib/spaced-recall/schedule.ts
type Grade = "again" | "hard" | "good" | "easy";

interface CardState {
  interval:   number;   // days
  ease:       number;   // starts at 2.50
  repetition: number;   // consecutive successful reviews
}

export function schedule(card: CardState, grade: Grade): CardState {
  if (grade === "again") {
    return {
      interval: 1,
      ease: Math.max(1.30, card.ease - 0.20),
      repetition: 0,
    };
  }
  const q = grade === "hard" ? 3 : grade === "good" ? 4 : 5;
  const ease = Math.max(
    1.30,
    card.ease + (0.10 - (5 - q) * (0.08 + (5 - q) * 0.02)),
  );
  const rep = card.repetition + 1;
  const interval =
    rep === 1 ? 1 :
    rep === 2 ? 6 :
    Math.round(card.interval * ease);
  return { interval, ease, repetition: rep };
}

card anatomy

Four fields. Each does one job.

A generated card is not a textbook flashcard — it’s a scaffolded reconstruction of the specific mistake you made, in the context of the specific problem where you made it.

Concept (front).
The concept, rephrased by Haiku so you can't win via surface memorization of the wording. If you failed the same concept twice, the second card asks about a different facet.
Attempted answer (back).
Your original attempt, in your own words, marked with the specific mistake — plus the corrected reasoning. Not a textbook answer; your answer, corrected.
Source pointer.
One-tap jump back to the original session where the error happened, so you can re-attempt the whole problem if the card alone isn't enough.
Schedule metadata.
SM-2 fields: interval, ease factor, repetition count, next due date. All queryable; nothing hidden.

from mistake to card

A wrong attempt becomes a card. In real time.

Below: the tutor session in which the mistake happened, followed by the card that appeared in the student’s recall queue an hour later.

AP Chemistry — limiting reagent

mistake → card

real transcript

student · confidence 4

2H₂ + O₂ → 2H₂O. Given 4g H₂ and 40g O₂, what's the limiting reagent? I think it's O₂ because there's more of it in grams.

brainback

Correction: 'more grams' ≠ 'more moles' ≠ 'more reactive units'. The mole ratio matters, not the mass. Convert both to moles first — what do you get?

student

4g H₂ = 2 mol. 40g O₂ = 1.25 mol. Oh — need mole ratio. 2 mol H₂ needs 1 mol O₂, so I have plenty of O₂. H₂ is limiting.

systemcard generated · concept: mass ≠ moles ≠ limiting reagent · scheduled tomorrow

One hour later, this card was in the student's queue: 'A student confuses grams with moles when identifying a limiting reagent. What's the correct sequence of conversions?' The wording is different from the original problem — the concept is exactly the one they failed.

Pre-made flashcard decks optimize for the average learner’s gaps. Yours are not average — they are yours. A card that tests knowledge you already have is a card that wastes your time. A card that skips knowledge you don’t have is a card that reinforces a hidden gap.

Brainback cards are generated from problems you got wrong and concepts you explained back poorly. They are rephrased by Haiku so you can’t rely on surface recognition — but the underlying concept is one you actually failed to demonstrate. Every review is a test of a gap you actually have.

SM-2, in plain English

SM-2 is a scheduling algorithm from 1990. It works. Every recall product worth using descends from it. Here’s how it applies inside Brainback:

New card: reviewed the next day.
Rated Good or better: next interval = previous × ease factor. Ease starts at 2.5.
Rated Again: ease drops by 0.20, interval resets to 1 day.
Rated Easy: ease climbs, interval multiplies by a larger factor.

Over weeks, cards you know well drift to longer intervals. Cards you struggle with come back often. The math does the tuning; you don’t.

The review UX

Cards flip on tap or spacebar. Before flipping, you’re asked for your confidence — one to five. That data feeds calibration. After flipping, four buttons: Again, Hard, Good, Easy. Any more granularity and users froze; any less and the scheduling got noisy.

The whole review is designed for one thing: speed. Median time to review a card in Brainback is nine seconds. You’ll do ten cards in less time than it takes to make coffee.

The offline story

The PWA caches your entire due queue on the last online visit. You can complete an entire review session on the subway. When you re-connect, grades sync back and the schedule updates. Because SM-2 is deterministic and the client can compute the next-due date locally, there’s no server round-trip during a review — the entire experience is native-app fast.

the case against pre-made decks

A pre-made deck of 2,500 First-Aid cards is a wall of noise. A generated deck of 60 cards, all built from your own mistakes over four weeks, is a signal. The compounding is bigger. The time-to-mastery is shorter. And the cards you cover are, by construction, the ones you needed.

fair questions

Everything students ask about recall.

Do I have to write my own cards?▾

No. Cards are generated automatically from your mistakes and low-scored explain-backs. You can write your own if you want (Student plan and up), but the ones the system writes on your behalf are, by design, the ones you actually needed.

How is this different from Anki?▾

Anki is a general-purpose SRS with a deck you assemble by hand or by download. Brainback is a domain-specific SRS with a deck that assembles itself from the errors you actually made in the tutor. The scheduling math is the same lineage; the card-generation half is completely different.

Can I import Anki decks?▾

Not currently. We may add limited import for high-value decks (First Aid for the USMLE, AnKing, etc.) if there's demand. The bar we set for import is that cards land in your deck without polluting the shape of what “your card” means.

What's the median time per card?▾

9 seconds. The confidence step + the flip + the four-button grade adds up to a fast review pass. Ten cards in ninety seconds is the target; that's a good study break, not a chore.

Does it work offline?▾

Yes. The PWA caches your due queue. Reviews complete offline and sync when you're back on network. Your interval math is deterministic — the queue you saw offline is exactly the queue you'll see after sync.

Can I export my cards?▾

Yes. Full JSON export from account settings. If you leave, you leave with your deck. This is written into the DPA and the account-deletion flow.

the rest of the engine

Every piece connects. Read the ones you haven’t.

Five rungs. Never the answer.

read

product

Explain-back

Teach it. Prove you know it.

read

product

Brain score

One number. Every week.

read

product

Efficiency mode

Normal AI for everything else.

read

product

Output guard

Every reply checked. No leaks.

read

→spaced recall

Ten cards, ninety seconds, ninety-two percent.

Free tier: 3 reviews a day. Student unlocks the whole queue.

The scheduling code, verbatim.

This is the entire scheduling function. Twenty-eight lines. Deterministic. Testable. If a card’s next-due date ever surprises you, this function is the answer.

lib/spaced-recall/schedule.ts

typescript

// lib/spaced-recall/schedule.ts
type Grade = "again" | "hard" | "good" | "easy";

interface CardState {
  interval:   number;   // days
  ease:       number;   // starts at 2.50
  repetition: number;   // consecutive successful reviews
}

export function schedule(card: CardState, grade: Grade): CardState {
  if (grade === "again") {
    return {
      interval: 1,
      ease: Math.max(1.30, card.ease - 0.20),
      repetition: 0,
    };
  }
  const q = grade === "hard" ? 3 : grade === "good" ? 4 : 5;
  const ease = Math.max(
    1.30,
    card.ease + (0.10 - (5 - q) * (0.08 + (5 - q) * 0.02)),
  );
  const rep = card.repetition + 1;
  const interval =
    rep === 1 ? 1 :
    rep === 2 ? 6 :
    Math.round(card.interval * ease);
  return { interval, ease, repetition: rep };
}

SM-2, in plain English

SM-2 is a scheduling algorithm from 1990. It works. Every recall product worth using descends from it. Here’s how it applies inside Brainback:

New card: reviewed the next day.

Rated Good or better: next interval = previous × ease factor. Ease starts at 2.5.

Rated Again: ease drops by 0.20, interval resets to 1 day.

Rated Easy: ease climbs, interval multiplies by a larger factor.

Over weeks, cards you know well drift to longer intervals. Cards you struggle with come back often. The math does the tuning; you don’t.

The review UX

The whole review is designed for one thing: speed. Median time to review a card in Brainback is nine seconds. You’ll do ten cards in less time than it takes to make coffee.

The offline story