Changelog for haskell-fsrs

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog. The major version tracks the FSRS algorithm version, as py-fsrs and fsrs-rs do: 7.x.y implements FSRS-7.

Unreleased

7.1.0 - 2026-08-21

Track the finished FSRS-7. Release 7.0.0 implemented the 35-weight, single-trace draft still published as models/fsrs_v7.py in srs-benchmark; the algorithm was subsequently finalised with 34 weights and a dual-trace memory model, and that is what "FSRS-7" now means. The reference is Expertium/fsrs-rs-speed-autoresearch, upstreamed into fsrs-rs in PR #426.

This is a breaking change to the model's numbers and to several signatures. Stored parameter vectors and stored memory states from 7.0.0 are not compatible.

Changed

• parameterCount is now 34, and defaultParameters, parameterBounds and orderingConstraints are the finished model's. The layout is: w0..w3 initial stability, w4..w6 difficulty, w7..w14 slow-trace stability, w15..w22 fast-trace stability, w23..w30 forgetting curve, w31..w33 the curve's modulation by difficulty and stability.
• MemoryState gained a third field, memoryStabilityFast: FSRS-7 tracks two memory traces. A new card's fast trace starts at fastTraceRatio (0.8) of its slow one.
• retrievability, retrievabilityDerivative and nextIntervalDays take a whole MemoryState instead of a Stability — the curve mixes both traces and is shaped by the difficulty too. So does nextReviewInterval.
• retrievability is rescaled into [1e-5, 1 - 1e-5], so retrievability p 0 is now just short of 1 rather than exactly 1.
• nextDifficulty takes the retrievability at review time: a lapse's step is weighted by how surprising the lapse was.
• stabilityAfterReview takes a stability and a difficulty rather than a MemoryState, since it now applies to whichever trace is being updated, and its post-lapse branch no longer depends on difficulty.
• nextStability returns both traces as a (slow, fast) pair.
• longTermWeights / shortTermWeights are now slowTraceWeights / fastTraceWeights, and StabilityWeights lost swFailureDifficultyExponent — the weight it read was ablated from both blocks.
• CurveWeights was reshaped: cwDecay1 became cwDecayBase1 (the fast component's decay is now scaled by the fast stability, so it is not a constant), cwDecay2 holds a magnitude rather than a negated value, and cwDifficultyWeight, cwDifficultyDecay and cwStabilityDecay1 are new.

Added

• fastTraceRetrievability — the fast trace's own recall probability, which is what drives its stability update.
• initialMemoryState — the state a card is born with.
• retrievabilityFloor and fastTraceRatio.
• orderingConstraints is now exported from FSRS.Parameters.
• The two numeric assertions the reference implementation makes about FSRS-7 in its own test suite are now reproduced in ours, in Haskell and in Python, so the port is pinned to upstream and not only to our own transcription.

Removed

• transitionCoefficient, transitionRate and transitionAmplitude, along with the two weights behind them. The elapsed-time blend between a long- and a short-term stability update is gone; the fast memory trace replaces it.

7.0.0 - 2026-08-20

Initial release: FSRS-7.

Added

• FSRS.Types — Rating, MemoryState and the Stability / Difficulty / Retrievability / Days synonyms.
• FSRS.Parameters — the 35 FSRS-7 weights (of the draft model, see 7.1.0), the default set, bounds taken from the upstream optimiser's clipper, validation with per-weight errors, clamping, and typed views onto the weight blocks (StabilityWeights, CurveWeights).
• FSRS.Algorithm — the model: the two-component retrievability curve and its derivative, nextIntervalDays (a safeguarded Newton root-find, since the FSRS-7 curve has no closed-form inverse), initialDifficulty, nextDifficulty, initialStability, stabilityAfterReview, transitionCoefficient, nextStability, nextMemoryState and replayReviews.
• FSRS.Scheduler — cards, due dates, learning and relearning steps, review logs, interval preview and explicit (pure) interval fuzzing.
• FSRS — an umbrella module re-exporting all of the above.
• A test suite of golden vectors generated from a pure-Python transcription of the upstream reference implementation, plus property tests for the model's invariants.