gaetan/knowledgify

Decision: SRS algorithm — SM-2, FSRS, or custom variant? #8

New issue

Closed

opened 2026-08-20 14:34:32 +02:00 by gaetan · 1 comment

gaetan commented

2026-08-20 14:34:32 +02:00

Owner

Part of #1

Question

Which spaced repetition algorithm should power the review scheduling?

The content model is settled (Q&A + Vrai/Faux, SM-2 fields in the DB), so we can now choose the algorithm.

Options:

SM-2 — The original Anki algorithm. Simple, well-understood, 5 rating buttons (Again, Hard, Good, Easy). Proven effective. ~50 lines of code.
FSRS (Free Spaced Repetition Scheduler) — Anki's newer ML-based algorithm. Better interval distribution, fewer reviews needed. More complex (~200+ lines), needs parameter training.
Custom simplified SM-2 — SM-2 with fewer ratings (e.g. 3 buttons: Wrong, Okay, Easy). Simpler UI but less granular scheduling.
SuperMemo 2 variant — SM-2 with adjustments (e.g. different ease factor decay, different interval multipliers).

Key considerations:

The rating mechanism (how the user signals recall quality) is tightly coupled to the algorithm
SM-2 uses 4 ratings (Again=1, Hard=2, Good=3, Easy=4) mapped to the algorithm's parameters
FSRS uses 4 ratings but interprets them differently (probability of recall)
A 3-button rating (Wrong/Okay/Right) is simpler but less precise

This decision shapes the review UI (rating buttons) and the scheduling logic.

Dependencies

Depends on: #4 Tech stack (framework choice affects implementation)

Part of #1 ## Question Which spaced repetition algorithm should power the review scheduling? The content model is settled (Q&A + Vrai/Faux, SM-2 fields in the DB), so we can now choose the algorithm. Options: - **SM-2** — The original Anki algorithm. Simple, well-understood, 5 rating buttons (Again, Hard, Good, Easy). Proven effective. ~50 lines of code. - **FSRS (Free Spaced Repetition Scheduler)** — Anki's newer ML-based algorithm. Better interval distribution, fewer reviews needed. More complex (~200+ lines), needs parameter training. - **Custom simplified SM-2** — SM-2 with fewer ratings (e.g. 3 buttons: Wrong, Okay, Easy). Simpler UI but less granular scheduling. - **SuperMemo 2 variant** — SM-2 with adjustments (e.g. different ease factor decay, different interval multipliers). Key considerations: - The rating mechanism (how the user signals recall quality) is tightly coupled to the algorithm - SM-2 uses 4 ratings (Again=1, Hard=2, Good=3, Easy=4) mapped to the algorithm's parameters - FSRS uses 4 ratings but interprets them differently (probability of recall) - A 3-button rating (Wrong/Okay/Right) is simpler but less precise This decision shapes the review UI (rating buttons) and the scheduling logic. ## Dependencies - Depends on: #4 Tech stack (framework choice affects implementation)

gaetan added the

wayfinder:grilling

label

2026-08-20 14:34:32 +02:00

gaetan added a new dependency

2026-08-20 14:34:37 +02:00

#4 Decision: Tech stack — Go web framework and frontend approach

gaetan referenced this issue

2026-08-20 14:35:59 +02:00

Decision: Card review UI — how the review screen works #11

gaetan added a new dependency

2026-08-20 14:36:07 +02:00

#11 Decision: Card review UI — how the review screen works

gaetan self-assigned this

2026-08-20 14:47:44 +02:00

gaetan commented

2026-08-20 14:49:45 +02:00

Author

Owner

Resolution

SRS algorithm decided:

Algorithm: SM-2 (Anki's original). ~50 lines of code. Proven, simple, adequate for 100-200 cards.

Rating buttons: 4 buttons (Again=red, Hard=orange, Good=blue, Easy=green) + keyboard shortcuts (1=Again, 2=Hard, 3=Good, 4=Easy). Review a card in under 3 seconds.

SM-2 parameters:

Initial interval: 1 day (after first "Good")
Interval cap: 30 days (for "Good" rating)
After 30 days: intervals grow by ease_factor * previous_interval
Ease factor: starts at 2.5, adjusted per answer
Lapse (Again on known card): reset repetitions to 0, interval to 0 (review again today)
New card limit: max 10 new cards per day

SM-2 formula:

new_interval = old_interval * ease_factor * adjustment
adjustment: Again=0, Hard=1.2, Good=2.5, Easy=4.0
ease_factor = ease_factor + (0.1 - (5 - rating) * (0.08 + (5 - rating) * 0.02))

Review flow:

Query cards where next_review <= now (due cards)
Prioritize: new cards first, then reviews by next_review ascending
Show card → reveal answer → rate → update progress → next card
When all due cards reviewed, show dashboard

## Resolution SRS algorithm decided: **Algorithm**: SM-2 (Anki's original). ~50 lines of code. Proven, simple, adequate for 100-200 cards. **Rating buttons**: 4 buttons (Again=red, Hard=orange, Good=blue, Easy=green) + keyboard shortcuts (1=Again, 2=Hard, 3=Good, 4=Easy). Review a card in under 3 seconds. **SM-2 parameters**: - Initial interval: 1 day (after first "Good") - Interval cap: 30 days (for "Good" rating) - After 30 days: intervals grow by `ease_factor * previous_interval` - Ease factor: starts at 2.5, adjusted per answer - Lapse (Again on known card): reset repetitions to 0, interval to 0 (review again today) - New card limit: max 10 new cards per day **SM-2 formula**: ``` new_interval = old_interval * ease_factor * adjustment adjustment: Again=0, Hard=1.2, Good=2.5, Easy=4.0 ease_factor = ease_factor + (0.1 - (5 - rating) * (0.08 + (5 - rating) * 0.02)) ``` **Review flow**: 1. Query cards where `next_review <= now` (due cards) 2. Prioritize: new cards first, then reviews by next_review ascending 3. Show card → reveal answer → rate → update progress → next card 4. When all due cards reviewed, show dashboard

gaetan removed a dependency

2026-08-20 14:50:20 +02:00

#4 Decision: Tech stack — Go web framework and frontend approach

gaetan closed this issue

2026-08-20 14:50:26 +02:00

gaetan referenced this issue

2026-08-20 14:51:00 +02:00

Wayfinder map: Anki-like spaced repetition web app for culture générale #1

gaetan referenced this issue

2026-08-20 15:23:59 +02:00

Decision: Spaced repetition algorithm — SM-2, FSRS, or custom? #3