SRS Algorithm: Comparison with Anki

Algorithm Overview

Our algorithm is a simplified variant of SuperMemo SM-2 (the foundation Anki is also built on). It uses a 4-point rating scale (Again/Hard/Good/Easy), a learning phase with fixed intervals, and an ease-factor-based review phase.

Core Implementation

See

DeckContext.tsx:320-334

for

calculateNextInterval()

and

DeckContext.tsx:220-294

for

updateCardProgress()

Design Philosophy

This app is built specifically for language learning, where the goal is not just correct recall but instant automaticity — the kind of reflex you need in a real conversation. This philosophy drives every design decision.

Hesitation = needs more practice. If you had to think, the word wouldn't have come in a conversation. "Good" (correct with difficulty) is treated as needing re-exposure, not as a pass.
Unified state model. Unlike Anki's learning/relearning/graduated stages, our algorithm uses one simple state. This is easier to understand and removes an unnecessary source of confusion for learners.
Stricter standard = more reviews. This is intentional. The tradeoff for building muscle memory is more frequent practice on cards that aren't yet automatic.

Comparison Table

Aspect	Our Algorithm	Anki
Ratings	0=Again, 1=Hard, 2=Good, 3=Easy (same scale)	0=Again, 1=Hard, 2=Good, 3=Easy
"Good" in review	Treated as failure → 180s	Success (interval grows, just less than Easy)
"Hard" in review	Treated as failure → 90s	Interval grows by 1.2x (no reset)
Learning steps	5 fixed (10min→1hr→4hr→1d→3d)	Configurable, default 2 (1min, 10min)
Ease on "Good"	Decreases by 0.10	No change
Ease on "Hard"	Decreases by 0.12	Decreases by 0.15
Ease on "Again"	Decreases by 0.14	Decreases by 0.20
Ease on "Easy"	Increases by 0.10	Increases by 0.15
Ease floor	1.3	1.3
Failure of mature card	Single fixed short interval (30s/90s/180s)	Goes through learning steps again, then re-graduates
Interval fuzzing	No	Yes (prevents card clustering)
Customizable steps	No (hardcoded)	Yes

Strengths (given our goals)

Simple & predictable — easy to explain to users and debug
Strict standard for fluency — "Good" is not a pass, matching real-world language use
Aggressive re-exposure — any hesitation means you see the card again very soon, building automaticity faster
Generous initial ladder — 5 exposures in the first day builds strong initial memory before graduation
No state complexity — no distinction between learning/relearning/graduated; one unified state model

Tradeoffs vs Anki

These are not weaknesses — they are intentional tradeoffs for the language-learning goal. Whether they matter depends on what the user wants:

Tradeoff	How it works	Rationale
"Good" resets you	Any non-perfect rating resets to sub-minute intervals, even for mature cards	In a conversation, hesitation = fail. You need automaticity, not just recall.
No "Hard" middle ground	Both "Hard" and "Good" are treated as needing re-exposure	The distinction between "wrong" and "right with effort" doesn't matter in real-time speech.
Ease drifts down	"Good" reduces ease, so consistent non-perfect performance shrinks intervals	This is the mechanism that ensures non-automatic cards get the reps they need.
No graduated relearn	One short interval for failure, then back to normal scheduling	Simpler model; rapid re-exposure is more important than a gentle ladder back up.
No interval fuzzing	Cards can cluster on the same day	Acceptable tradeoff for simplicity; the app limits total cards per session anyway.
No customization	Learning steps are hardcoded	Reduces complexity for the target audience (language learners, not power users).

When to use which

Use our algorithm if you're learning a language and want conversational automaticity, not just recognition. You accept more reviews in exchange for faster reflex building.
Use Anki if you want maximum efficiency per review, prefer fine-grained control, or are studying subjects where "correct with effort" is a meaningful success (e.g., medical school, trivia).