From a document to something you still know in six months

Everything starts with a space, which is one subject: a certification, a course module, a language, a talk. Keeping subjects separate matters more than it sounds, because finance, law and operations reuse the same words for different things, and a shared pool would schedule one subject's term against another's definition.

A space can be shared with other people as owner, editor or viewer through a single-use invite link, so a study group or a team maintains one item bank rather than several overlapping ones.

There are two ways to fill a space. The first is to supply material: paste text, upload a document, or point at a web page. The document is split into chunks, embedded, and deduplicated against what the space already holds, so re-uploading a revised edition does not double every card.

The second is to describe a topic and let the app draft cards for it with no document behind them. This is genuinely useful when you have no source at hand, and it is the case the headline claim is built around: those cards carry no citation, and they are labelled as unsourced rather than being dressed up as grounded.

For sourced material, each generated card stores the exact span it was drawn from. Before the card is accepted, that span is validated as a literal substring of the ingested chunk. Exact matching is deliberately strict, because the failure it catches, a quote that was paraphrased or invented during generation, is the failure that matters.

Matches that are ambiguous, typically where whitespace or formatting differs, escalate to an entailment check instead of being waved through. Cards that fail both are not shipped; they land on the card quality report where you can see why.

Generation runs as tracked jobs with retries and idempotent unwinding, so a failure part-way through a large document does not leave half a space behind.

Reviews are scheduled by FSRS-6, the sixth version of the Free Spaced Repetition Scheduler, through the ts-fsrs package. FSRS models a forgetting curve per card from your own answer history and schedules the next review shortly before predicted forgetting, rather than stepping through fixed intervals the way SM-2 does.

The practical effect is that a fixed daily budget covers more material: cards you find easy stop consuming review slots quickly, and the difficult middle band gets the attention.

Generous self-grading is the standard failure of self-directed spaced repetition, because the schedule is only as good as the grades feeding it. Concept mode closes that hole in a specific way: you write your recall and commit it, and the model answer, the rationale and the anchor term are not present in the page at all until you do. There is nothing to peek at.

After you commit, the app checks that the anchor term genuinely appears in what you wrote and asks a short binary question about it. A blank or off-target answer cannot be graded correct.

Verbatim mode handles material that must be exact. The text fades progressively across passes until you can produce it unaided, and progress is recorded per passage, so the paragraph that keeps collapsing gets the practice.

Mastery has a defined bar: three consecutive correct recalls, a retained interval of at least three weeks, and at least one confident answer. A single correct answer never counts, and neither does a run of them at short intervals.

Two signals are tracked separately from correctness. Confidence is what you claimed before seeing the answer, and hesitation is how long you took. Answers you were sure about and got wrong are surfaced first, because those are the beliefs that survive into the exam room unexamined.

Progress, quality and exam views share one filter vocabulary, Not started, In progress and Completed, and the filters persist per space so a view reopens exactly as you left it.