Never lose a reference.

Spaced repetition

The gap between reviews matters more than the hours you put in.

Six weeks of notes and almost none of it stuck, which is the opposite of how I had assumed revision was supposed to work. The point was never to log more time at the desk. It was to be asked the question again just as the answer started to slip, which is a different thing entirely and much harder to schedule by feel.

What the studies agree on

• Retrieval beats re-reading, even when it feels harder in the moment
• The optimal gap widens after each successful recall of the same item
• Spacing related items apart cuts the interference between them
• Testing yourself is not just measurement, it is itself the study

Every scheduler still guesses at the timing. SM-2 quietly underpins most of them, and it was tuned back in 1987 on one person's flashcards. Nothing since has really replaced it, so the defaults every app ships with today are still, in effect, that same person's, carried forward because they were there, not because anyone re-ran the experiment at scale.

Conditions that slow acquisition often improve long-term retention (Bjork, 1994).

So I am rebuilding the deck around the gaps rather than the totals: shorter sessions, wider spacing, and no more re-reading the same page the night before.

Where it breaks down in practice

None of this is an argument against the algorithm, only against trusting it past the point where it is measuring anything real. A scheduler knows when a card was last seen and whether the answer came back right. It does not know that the answer came back right because the card sat directly underneath one that gave it away, or that two cards in the same deck are quietly teaching the same fact in slightly different words.

The failure mode is always the same, and it is not the maths. A deck grows past the point where its intervals mean anything, the daily queue turns into a backlog, and the backlog gets cleared by pressing the middle button on everything in it, which the scheduler reads as evidence that the intervals were correct, and duly widens them. The numbers stay healthy the whole way down.

• Leeches: the few cards that fail over and over and eat a third of the session
• Interference: near-identical cards that trade places in memory week to week
• Backlog: a queue long enough that the grading stops being honest

What I have settled on is duller than a better algorithm and works far better than one. Delete the leeches instead of nursing them. Write cards that ask for one thing, so a wrong answer means something specific rather than that the card was too broad to answer. Keep the deck small enough that a bad week can be caught up in a day, because the moment it cannot be, every interval in it is a guess resting on a guess.