From Leitner to FSRS

Judgement

Scheduler history is a progression in responsiveness, not a ladder of proven superiority. Physical compartments made review priority visible. Computerised rules calculated per-item intervals. Fitted models learned parameters from review data. Each design added machinery for choosing reviews, but newer machinery should not inherit the certainty of the spacing and retrieval findings beneath it.

The diagram traces added scheduling machinery from physical compartments through fixed per-item rules to fitted models, without treating progression as proof of superior retention.

flowchart TB
    A["Physical boxes<br/>Visible review groups<br/>Procedure known here second-hand"]
    B["Fixed formulae<br/>Per-item rules from ratings<br/>Added automatic timing"]
    C["Fitted models<br/>Parameters learned from logs<br/>Added data-driven prediction"]
    A --> B --> C

Figure: Scheduling became more responsive and data-driven, but the progression alone does not establish better long-term retention. Secondary accounts attribute the Leitner system to a 1972 book and describe successful cards moving toward less frequent compartments while failed cards return toward frequent review. The original book was not retrieved for this dossier, so exact box arrangements and procedures remain second-hand.¹

Later documentation describes SM-2 as a computer scheduler with a separate easiness factor for each item. It updated that factor from a learner-entered response grade and used it to calculate later intervals.² This replaced manual movement between compartments with automatic, item-sensitive rules. The documentation is a practitioner source, not a comparative trial.

Fitted models added learning from accumulated data. Duolingo's Half-Life Regression used learner performance and linguistic features to estimate memory for language material. It improved recall prediction against the paper's baselines, while its operational result measured engagement rather than an independent long-term retention test.³ Current FSRS documentation describes a model fitted to review histories.⁴ Interval choices are also shaped by a retention target the user sets in the scheduler.⁵ Anki now offers it as an alternative to a legacy SM-2-derived scheduler.⁵

With a deck and twenty minutes, a fitted scheduler can use your history instead of applying the same fixed rules indefinitely. That is a practical capability, not proof that it protects knowledge better. Choose a tool whose ratings you can use consistently and whose workload you can sustain. Do not switch merely because a model has more parameters or a newer name.

Boundary

Automated scheduling is a reasonable procedure built on the solid effects of spacing and retrieval. The algorithmic progression is not itself evidence that each generation produces better external long-term retention. The dossier contains no randomised comparison of current FSRS and SM-2 in ordinary Anki users using an independently administered retention test.