Methods — the science Sapior is built on
The science Sapior is built on
Sapior implements three findings from learning research that decades of replication keep validating. Here's what each finding says, the studies behind it, and how the engine implements it.
Finding 1 — Retrieval practice
Pulling a fact out of memory builds memory more than re-reading it does. This is one of the most replicated findings in learning research.
Adesope, Trevisan, and Sundararajan (2017) reviewed 272 testing studies in a meta-analysis published in Review of Educational Research. The average effect was g = 0.70, rising to g = 0.82 at one-to-six-day delays. Effects held across age groups, subject matter, and question format.
How Sapior implements this: every session is structured as retrieval — Sapior asks the AI to test, not re-explain, unless the material is genuinely new. Brand-new items get taught first; everything else is recall.
Finding 2 — Spaced practice
How long you wait between reviews of a fact changes whether you remember it. The optimal gap depends on how long you need to retain the material, and it shifts as your retention strength grows.
Cepeda et al. (2008), Psychological Science, mapped this relationship across 26 spacing conditions (N = 1,354) and fit a curve with R² = .98 to the resulting retention surface. Rawson and Dunlosky (2013) showed that when retrieval practice and spacing are combined — what they called successive relearning — the effect compounds: d = 2.37 on delayed recall at 24 days.
How Sapior implements this: the scheduler is built on FSRS v6 (Free Spaced Repetition Scheduler, version 6), the modern open-source algorithm derived from the Difficulty-Stability-Retrievability model. FSRS v6 tracks the retention curve for each item per learner and surfaces review at the moment retention is starting to decay. It's the open-source engine that today's modern spaced-repetition apps are built on. Implementation: ts-fsrs (MIT), maintained by Open Spaced Repetition.
Finding 3 — Expertise reversal
Brand-new material needs explanation. Once it's stuck, explanation actively gets in the way of retrieval — the same scaffolding that helps a novice slows down an expert.
Kalyuga (2007), Educational Psychology Review, formalized this as the expertise-reversal effect across a body of cognitive-load studies: instructional supports that benefit learners with low prior knowledge can harm learners with high prior knowledge on the same material. Supports should fade as mastery grows.
How Sapior implements this: the scheduler routes each interaction to one of three content phases based on the learner's current retention level for that item — teach (explain + worked example) for new material, review (guided retrieval with hints available) for partially-retained material, assess (pure recall, no hints) for well-retained material. Scaffolding shrinks automatically as retention rises.
What's in the engine vs. what's still ahead
Some of the science is implemented. Some is on the roadmap. To be precise about what runs today:
FSRS v6 with default parameters — runs for every learner. The retention model is per-item: each item you've studied has its own retention/stability/difficulty state.
Content-phase routing by retrievability — implemented. The scheduler decides between teaching, guided review, and pure recall.
Per-learner FSRS weight fitting — the database schema is ready, but the optimizer that fits weights from accumulated review history is not yet wired up. Every learner currently runs FSRS defaults. The infrastructure to back-fit per-learner weights is in place; the fitting routine is planned.
This page will be updated as the engine implementation expands.