Spaced retrieval for code: FSRS against 18,000 review events

The most replicated finding in cognitive science is also one of the least intuitive: to make something stick, retrieve it rather than re-read it. Testing yourself beats reviewing notes, and it is not close. Spaced repetition operationalises this by scheduling each retrieval just before you would have forgotten.

All of that literature is built on flashcards. An item is atomic, a lapse costs you one word, and success is binary. A code concept is none of those things. This paper is about what survives the move.

Three schedulers

We compared FSRS-5 against SM-2, the algorithm most spaced-repetition software used for decades, and against a fixed ladder of 1, 3, 8, 21 and 60 days. Every arm used identical review content and identical grading.

FSRS needed 31% fewer reviews than SM-2 and 54% fewer than the fixed ladder to hold the same predicted retention. That is a real efficiency gain and it matters: every review is a claim on the attention of someone who is trying to do their job.

Predicted retention is not measured retention

A scheduler optimises predicted retention using its own model of forgetting. The number that matters is whether the engineer can still demonstrate the concept when we check independently. Ours reached 71% at 30 days against a 75% target, which is roughly where well-calibrated Anki users sit on well-curated decks, and below where we want to be.

The aggregate hides the finding. Syntactic and API-level concepts retained at 84%. Architectural concepts, the kind where understanding means holding a system in your head, retained at 58%. A 26 point spread inside one average is a sign that the average should not be reported alone, which is why the product shows the breakdown.

Two modifications the domain forced on us

Prerequisite-aware scheduling

FSRS treats items as independent. Code concepts form a directed graph: you cannot meaningfully retrieve transaction isolation while your grasp of atomicity has lapsed. Scheduling them independently produces reviews the engineer fails for reasons that have nothing to do with the concept under test, which then corrupts the difficulty estimate for that concept.

We modified the scheduler to suppress a review when any prerequisite is below a stability threshold, and to schedule the prerequisite instead. This reduced spurious lapses by 22% and improved measured retention on dependent concepts by 6 points.

Depth-proportional lapse penalty

In vocabulary, a lapse costs one word. Forgetting an architectural concept invalidates everything built on it. Standard FSRS applies a uniform penalty to the stability estimate on failure; we scale it by the concept's depth in the prerequisite graph, so a foundational lapse resets more aggressively than a leaf one.

The modifications cost us 0.4 reviews per concept and bought 5 points of measured retention. Note that the shipped configuration is slightly less efficient than unmodified FSRS on the metric FSRS optimises, and better on the metric we care about. That is the whole argument for measuring retention independently of the scheduler.

The grading problem underneath all of this

Every scheduler depends on an honest recall grade. Anki asks the user to grade themselves, which works because the user wants to learn Japanese. An engineer under sprint pressure grading their own comprehension has a different incentive.

We grade from the explain-check rubric rather than self-report. Comparing the two on 4,200 events where we had both, self-reported grades were higher than rubric grades 68% of the time, with a mean inflation of 0.7 points on a four-point scale. A scheduler fed self-reports would have concluded that retention was 12 points better than it was.

Limits

• 71% against a 75% target is a miss, and the architectural subcategory at 58% is the part we would not defend as adequate.
• Five months is short relative to the 60-day intervals the scheduler assigns, so long-interval behaviour is extrapolated rather than observed.
• The prerequisite graph is hand-curated for 3,410 concepts and is certainly wrong in places. Errors there propagate into scheduling.
• Concept categories were assigned by us, and the syntactic/architectural boundary is fuzzier than a two-way split implies.
• FSRS-5 parameters were fitted on our own event log, so the comparison against SM-2 flatters FSRS by giving it a domain fit SM-2 never had.

Fitted parameters, the prerequisite graph, the anonymised review event log and the grading comparison subset are published.

Related work

The forgetting curve this work schedules against is Ebbinghaus [1], and the spacing effect is surveyed comprehensively by Cepeda and colleagues [2]. We use FSRS [3] rather than SM-2 [4] for the reasons its authors document, and our efficiency comparison reproduces the direction of their result on a very different item type. The prerequisite-aware modification is motivated by knowledge-space theory [5], which models exactly the dependency structure that flat scheduling ignores. Our grading concern is the metacognitive-accuracy problem described by Dunlosky and Lipko [6]: learners systematically overestimate what they have retained, which makes self-reported recall grades unsafe input to a scheduler. The comprehension instrument used for grading is documented in AWL-2026-06 [7], and the tagger that identifies concepts in AWL-2026-03 [8].

Artefacts

Everything below is published or available on request. A number nobody can reproduce is an advertisement, not a result.

References

1. Ebbinghaus, H. (1885). Uber das Gedachtnis. Duncker & Humblot.
2. Cepeda, N. J., Pashler, H., Vul, E., Wixted, J. T., & Rohrer, D. (2006). Distributed Practice in Verbal Recall Tasks. Psychological Bulletin, 132(3).
3. Ye, J., Su, J., & Cao, Y. (2022). A Stochastic Shortest Path Algorithm for Optimizing Spaced Repetition Scheduling. KDD.
4. Wozniak, P. A., & Gorzelanczyk, E. J. (1994). Optimization of Repetition Spacing in the Practice of Learning. Acta Neurobiologiae Experimentalis, 54.
5. Doignon, J.-P., & Falmagne, J.-C. (1999). Knowledge Spaces. Springer.
6. Dunlosky, J., & Lipko, A. R. (2007). Metacomprehension: A Brief History and How to Improve Its Accuracy. Current Directions in Psychological Science, 16(4).
7. Bhardwaj, M. (2026). The comprehension gap: 200 engineers, 12 months, and what agentic coding did to recall. AI Work Lab Research, AWL-2026-06.
8. Bhardwaj, M. (2026). A concept tagger that does not lie: calibration across 41 repositories. AI Work Lab Research, AWL-2026-03.