README

A Moodle activity module that schedules spaced repetition of question bank items, so learners revisit each question just as they are about to forget it.

Unlike Anki-style tools, learners never grade their own recall. The scheduling signal is derived entirely from whether the submitted answer was correct, as judged by Moodle's own question grading. That removes self-assessment bias, which is unreliable in classroom populations and impossible to enforce in assessed contexts, and it works with real question types — multiple choice, short answer, matching, cloze — because grading and rendering are done by the core question engine rather than reimplemented.

Requirements

Moodle 5.2 ($plugin->requires = 2026042000, supported branch 502)
PHP 8.2+

The plugin depends on 5.x-shaped question bank APIs: question banks as mod_qbank instances, and question_bank_entries as the stable identity of a question. It will not run on 4.x.

Installation

Copy the plugin into mod/rememberme in your Moodle tree (under public/ on 5.0+), then visit Site administration → Notifications to install.

How it works

The scheduling model

The scheduler follows FSRS-style memory modelling rather than SM-2. It stores two latent variables per learner per question — stability and difficulty — and derives the interval from stability at review time.

The distinction matters in practice. SM-2 stores an interval and an ease factor, and a lapse typically resets the interval to day one. Here a lapse reduces stability sharply but does not discard it, so a long-known item that slips once returns sooner than a brand new one instead of starting from scratch. Over a course where learners see the same bank all term, that difference is the difference between a manageable queue and a punishing one.

The interval is never stored, only derived, so re-tuning the model does not leave stale intervals behind in the database.

Correctness becomes a rating

The memory model expects a rating of again, hard, good or easy. That is synthesised from the objective grade:

Condition	Rating
Wrong, or partial credit below the threshold	`again`
Partial credit at or above the threshold	`hard`
Fully correct, slower than the learner's usual pace	`good`
Fully correct, at or faster than their usual pace	`easy`

Answer speed is a secondary signal only. It can never turn a correct answer into a lapse — it only separates good from easy, both of which lengthen the interval. It is compared against the learner's own rolling median for that question type, needs at least eight samples before it is trusted at all, and is discarded when the attempt was left open too long. With it switched off the mapping collapses to wrong/right, and the system works correctly in that mode: the mapper is written binary-first, so the simple path is the tested one.

The mapping is a replaceable strategy class, because different courses want different thresholds and researchers will want to swap it out entirely.

Tiered question pools

The pool is not flat. A teacher binds question categories in an ordered sequence, and new items are drawn from the learner's current band and every band below it — so a category left unfinished before moving on is not stranded. A band may draw on several categories, and the settings form offers a question bank to pick them from. Unlocking is per learner, so two students in one course can be at different points. Three modes:

Coverage (default) — the next band unlocks once every question in the current band has been seen at least once, whatever the learner made of them. It is the default because it asks only that the syllabus has been covered: a timer moves a learner on whether or not they met the band, and mastery can hold them behind a handful of items they keep lapsing.
Time — one band per interval, counted from the learner's first session rather than course start, so somebody who joins in week three is not handed four bands at once.
Mastery — the next band unlocks when a configurable proportion of the current band reaches a stability floor. Unseen items count against the threshold, so a band cannot qualify until most of it has been attempted. A backstop advances a stalled learner anyway after a maximum time, and flags that it did so: a learner who never leaves band one is the one who most needs the coverage.

Bands gate introduction, never revision. Once seen, an item competes for review on memory strength alone, whichever band it came from.

Grading is adherence, not accuracy

Grading accuracy would contaminate the very signal the scheduler depends on, by rewarding guess avoidance and answer lookup. So the grade measures whether learners keep up with their schedule.

Each week the learner must clear what is due. Two rules make that fair:

The target is frozen when the week begins. It is what was due at that moment plus the new items they may draw. Items that come due again during the week roll into next week instead of enlarging this one. Under the alternative — a denominator that grows as answered items fall due again — a learner who answers everything asked of them can never reach 100%, because every answer breeds another review before the week is out. The finish line recedes as they approach it. The test suite asserts this directly.
Partial weeks earn partial credit, capped at 1.0, so a learner who logs in late and does what they can is not scored as harshly as one who never came.

Returning on time is rewarded. Answering an item close to when it falls due earns grace, up to a maximum the teacher sets. Punctuality is measured rather than visits, because opening the activity is free and counting visits would reward the appearance of the habit rather than the habit. A learner who saves everything for one sitting a fortnight earns none of it, because their questions sat overdue. It is paid in grace, so it can only repair a bad week.

Getting a question wrong costs time, never marks. A wrong answer brings the question back within the same sitting rather than the next day. Weekly credit counts distinct questions engaged with, so answering one question repeatedly earns one point, and an answer submitted faster than the question could be read earns nothing at all.

Grace credit is a pool of fractional credit, not a count of whole weeks. It tops a week up toward 1.0 and costs exactly the gap it fills: rescuing a missed week costs 1.0, patching a 0.9 week costs 0.1. It is allocated at the end of the course, cheapest gaps first, so nothing is wasted early on a week the learner would have absorbed anyway.

Progress is shown as a personal streak. There is deliberately no leaderboard: because the queue is capped and driven by each learner's own memory state, the learner with the most reviews is the one with the most lapses, so a leaderboard would rank learners roughly inversely to how well they know the material.

Suspension windows

A teacher can declare breaks during which the scheduling clock stops. This is the part that is easy to get wrong. If scheduling ticked through a two-week break, everything would fall due at once and learners would return to a wall of overdue reviews created purely by a holiday somebody else declared — punishing them for the break and corrupting the difficulty estimates for the whole cohort.

Rather than batch-shifting stored dates when a window ends (which breaks for mid-window enrolments, cannot be undone if the teacher edits the window, and corrupts the review log's elapsed times), suspended time simply does not exist. Every scheduling calculation runs through an effective-time function, so windows stay editable after the fact and are correct for learners who join during one. Weeks more than half suspended drop out of the grade entirely.

Voluntary study during a break earns grace credit, capped so a break cannot be farmed to buy back an absent term. At most one band unlocks per window, so a motivated learner does not mortgage their first week back.

The learner experience

Opening the activity is the session — no landing page, no start button. The friction of an intermediate screen is disproportionate for something meant to be visited briefly and often. Answer, get immediate feedback with an optional audio cue, and the next question appears, all without a page reload. The pause is longer after a wrong answer, because there is more to take in.

Multiple choice questions are presented as options the learner taps, with no submit button: the tap is the answer. The letter on each option turns into a tick or a cross in place, so the result never rests on colour alone. There is no confirmation step, which is reasonable here because being wrong costs a repetition and never a mark.

Answer options are shuffled every time, whatever the question was authored to do, and a teacher can cap how many options a question presents — a question written with eight options is a reading exercise on a phone. The right answers are always kept and the wrong ones thinned at random, drawn again each time the question comes round, so the shape of the answer cannot be memorised in place of the answer.

Scheduling state is written per question as it is answered, so a learner who closes the tab after three questions keeps the effect of those three.

Teacher reports

Question difficulty across the cohort — surfaces defective items, since a question that is hard for nearly everyone is usually badly worded rather than conceptually difficult.
Coverage and retention per learner, with a review-load forecast.
Band progression, flagging learners advanced by the backstop.
Weekly completion matrix, including grace consumed.

Mobile

The Moodle app shows the study session itself, not a summary of it. Questions are rendered by the app's own question component, so every question type the app understands works without this plugin knowing anything about them, and multiple choice gets the same tappable options as the web view. The app view has been exercised in a real Moodle app, though not yet on a physical device.

The site has to be set up for the app before any of this works — web services on, the REST protocol enabled, and the webservice/rest:use capability granted, which the admin UI's mobile setting does and the CLI equivalents do not.

The app assumes a live connection. Offline study is deliberately out of scope for this release: it requires queueing attempts locally and replaying them, which raises questions this design does not yet answer — what counts as due while disconnected, how to reconcile a replayed attempt whose elapsed time is now stale, and what happens when the same question is answered on two devices before either syncs.

Privacy

The schedule and the review log record what an individual learner knew and when. Both are declared, exported and deleted through the privacy API, along with weekly progress, band progress and session records.

Status

Version 0.1.1, alpha. The model constants are the published FSRS-5 defaults and the Mode B thresholds are reasoned estimates rather than validated ones — the review log exists precisely so they can be refitted against real cohort data.

Licence

GNU GPL v3 or later.

adamjenkins / moodle-mod_rememberme

Maintainers

Package info

Statistics

Security