How EdenMind works

Each piece is written to be read in under ten minutes, in short passages with figures, and most also have an audio episode of about the same length.

A piece opens with a question the reader is not expected to answer yet. Attempting a question before studying the answer improves later memory for that answer, even when the attempt is wrong; this is called the pretesting effect.³ Questions follow the passages that answer them, and the piece ends with a recap to be recalled with the text out of sight.

Adding a piece to your library puts its questions in the app on your phone. Each answer is a word or a short phrase, and you say it aloud. A session of twenty questions takes a few minutes.

Each answer is an act of retrieval. In one experiment, students tested five minutes after studying remembered slightly more of a passage they had reread than of one they had recalled. Tested a week later, they remembered more of the recalled passage, 56% against 42%.⁴ In a second experiment, the students who had only reread were the most confident that they would remember it.

Answering aloud has a smaller benefit of its own, and a less certain one. Words read aloud are remembered better than words read silently, which is called the production effect.⁵ For single words, though, recalling silently has been found to help memory as much as answering overtly⁶; an advantage for producing the answer has been shown with longer material, such as definitions.⁷ The main reason the app asks for a spoken answer is that it can be checked.

In the usual flashcard method, the learner sees the answer and then judges whether they knew it. That judgement is unreliable: with the answer in view, people overestimate what they would have recalled without it.⁸ In one study, students who scored their own answers too generously stopped practising those items too early, and remembered less of them later.⁹

In EdenMind the answer is marked before you see it. A speech recognition model transcribes what you said, and the transcript is compared with the expected answer by rules written for speech: they allow for an accent, a plural, a number said in words and a word misheard by a letter or two. Longer answers are compared by meaning, using a small language model.

A clearly correct answer is accepted and the app moves on; you can undo it if it misheard you. A correct answer given quickly is scheduled further out. Anything else shows the expected answer and any explanation, what the app heard and a suggested grade, which you can accept or change.

New questions show a hint, which stops appearing once you know the answer.

Each question has its own schedule, computed from your answers to it. A question you know comes back after days, then weeks, then months; one you miss comes back later in the same session and again soon after.

Spreading reviews out like this is the distributed practice described above; a 2006 review analysed 317 experiments on it.¹⁰

The scheduler is FSRS, an open-source model also used in Anki.¹¹ For each question it estimates the probability that you can still recall the answer, and schedules the next review for when that probability is predicted to fall to 90%. The figure is a model’s prediction, not a measurement, and you can change the target in Settings.

New questions are introduced a few at a time, at a pace you set. A reminder is sent on days when reviews are due, and not otherwise.

Two pieces cover this research in more detail: The testing effect and Spaced repetition.

Progress shows how your recall is holding, which questions are slipping and what is due over the next week. The streak counts the days on which reviews were due, so a day with nothing due does not break it.

It also projects how your recall would decline over the next month if you stopped reviewing.

A map links questions that share a concept across Topics, for example where the same idea appears in physics, chemistry and biology. Each question is a point, coloured by how well you hold it.

The pieces cover mathematics, physics, chemistry, biology, psychology and society. They are written in plain language for adults who are curious but out of practice, with Muslim readers in mind, and they assume nothing from school.