ADR 0005 Scheduling

ADR-0005: SM-2 scheduling over an append-only review log

Source of truth: this wiki page.

• Status: Accepted. Superseded in part by ADR-0033: FSRS replaces SM-2; the review log and per-pair scheduling stand. Amended by ADR-0034: which pairs are scheduled.
• Date: 2026-09-19

Context

Spaced repetition needs two things: an algorithm deciding when a card is next due, and somewhere to keep the state it operates on.

Decision

Implement SM-2. Persist every review event to an append-only log, and treat per-card scheduling state as a derived cache.

Schedule per (card, drill mode), not per card.

Consequences

The log is the important half of this decision. Storing only the current interval — which is what most implementations do — permanently discards the data needed to evaluate or change the algorithm. With a full log, FSRS or anything else can be swapped in later and the entire history replayed through it. Statistics that were never anticipated remain computable.

The cost is storage, and it is negligible: a review record is well under 200 bytes, so a decade of heavy daily use stays in single-digit megabytes.

Per-mode scheduling reflects how language learning actually works. Recognising a word is easier than producing it, which is easier than recognising it by ear. One shared interval would over-drill the easy direction and under-drill the hard one. The cost is up to four scheduling rows per card.

SM-2 over FSRS because it is ~100 lines with no dependency, is well understood, and is good enough. FSRS gives better retention per review, and the log means adopting it later loses nothing.

SM-2 over Leitner because Leitner's fixed buckets ignore how hard a card is for the individual learner.

The scheduler is pure Dart with no Flutter dependency, so it is unit-testable without a device.

See also: Contributing · ADR-Index · Home