@squeakyrobot/fsrs

A pure TypeScript implementation of the Free Spaced Repetition Scheduler (FSRS) v4.5 algorithm with optional v6 support.

Overview

FSRS is a modern, evidence-based spaced repetition algorithm that optimizes review intervals based on the DSR (Difficulty, Stability, Retrievability) model of memory. Unlike traditional algorithms like SM-2, FSRS uses trainable parameters optimized on real user data to achieve superior retention prediction.

Key Features

• FSRS v4.5 Core: Proven algorithm with stable defaults (17 parameters)
• 81% Better Than SM-2: Captures 98% of v6's improvement with fewer parameters
• Edge Runtime Ready: Works in Cloudflare Workers, Vercel Edge, Deno Deploy
• Continuous Grading: Supports both discrete (1-4) and continuous (1.0-4.0) ratings
• Auto-Rating: Built-in response time to grade conversion
• Optional v6 Support: Accepts optimized 21-parameter sets for advanced users
• Type-Safe: Full TypeScript support with strict typing
• Zero Dependencies: Minimal bundle size
• Pure Functions: Immutable API, no side effects

Installation

npm install @squeakyrobot/fsrs

Quick Start

Basic Usage with Discrete Ratings

import { FSRS, Rating } from '@squeakyrobot/fsrs';

// Initialize with default v4.5 parameters
const fsrs = new FSRS();

// Create a new card
let card = fsrs.createEmptyCard();

// Simulate a review session - get all 4 possible outcomes
const now = new Date();
const reviewResults = fsrs.repeat(card, now);

// User rates the card as "Good"
const { card: updatedCard, log } = reviewResults[Rating.Good];

console.log(`Next review in ${updatedCard.scheduled_days} days`);
console.log(`Current stability: ${updatedCard.stability} days`);
console.log(`Difficulty: ${updatedCard.difficulty}/10`);

Auto-Rating with Continuous Grades

import { FSRS } from '@squeakyrobot/fsrs';

const fsrs = new FSRS();
let card = fsrs.createEmptyCard();

// Track response time
const start = Date.now();
// ... user answers flashcard ...
const responseTime = Date.now() - start;

// Convert response time to grade (1.0-4.0)
const averageTime = 2000; // Expected average response time (ms)
const grade = fsrs.autoRating(responseTime, averageTime, card.difficulty);

// Schedule with continuous grade
const { card: updatedCard, log } = fsrs.scheduleWithGrade(card, grade);

Edge Runtime Example (Cloudflare Workers)

// worker.ts
import { FSRS } from '@squeakyrobot/fsrs';

const fsrs = new FSRS();

export default {
  async fetch(request: Request, env: Env): Promise<Response> {
    const { card, responseTime } = await request.json();

    const grade = fsrs.autoRating(responseTime, 2000);
    const result = fsrs.scheduleWithGrade(card, grade);

    return Response.json(result);
  }
};

Algorithm Background

For complete mathematical details, see the Algorithm Reference.

The DSR Model

FSRS models memory using three distinct variables:

1. Difficulty (D): Inherent complexity of the item (1.0 = easiest, 10.0 = hardest)
2. Stability (S): Time interval (days) for retrievability to decay from 100% to 90%
3. Retrievability (R): Instantaneous probability of successful recall (0-1)

Why V4.5?

V4.5 provides the optimal balance for a minimal implementation:

• Proven defaults: Works excellently without requiring parameter optimization
• 98% of v6's benefit: Only 3% RMSE difference from v6 on benchmark data
• Simpler: No same-day review complexity
• Stable: Widely deployed with deterministic behavior

V6 support is available for advanced users who:

• Run the FSRS optimizer on their review history
• Need personalized parameters (w0-w20)
• Want the marginal 2-3% accuracy improvement

API Reference

For complete API documentation, see the Full API Reference.

FSRS Class

Constructor

constructor(params?: Partial<FSRSParameters>)

Creates a new FSRS scheduler instance with optional parameter overrides.

Methods

repeat(card: Card, now?: Date): SchedulingCards

Calculates scheduling outcomes for all possible ratings. Useful for showing users all 4 button options with predicted intervals.

Parameters:

• card: Current card state
• now: Current timestamp (default: current time)

Returns: Map of Rating -> { log, card } for all 4 ratings

scheduleWithGrade(card: Card, grade: Rating | ContinuousRating, now?: Date): SchedulingResult

Schedules a review with a specific grade. Primary method for auto-rating scenarios and manual reviews.

Parameters:

• card: Current card state
• grade: Rating (discrete 1-4) or continuous (1.0-4.0)
• now: Current timestamp (default: current time)

Returns: { log, card } with updated card state and review log

autoRating(responseTime: number, averageTime: number, difficulty?: number): ContinuousRating

Auto-rating utility that converts response time to a continuous grade.

Parameters:

• responseTime: Time taken to respond (milliseconds)
• averageTime: Expected average response time (milliseconds)
• difficulty: Optional card difficulty (1-10) for adjustment (default: 5.5)

Returns: Continuous rating (1.0-4.0)

Formula:

• Much faster than average: approaches 4.0 (Easy)
• Around average: approaches 3.0 (Good)
• Slower than average: approaches 2.0 (Hard) or 1.0 (Again)

getRetrievability(card: Card, now?: Date): number

Calculates current retrievability (probability of recall).

Parameters:

• card: Card to evaluate
• now: Current timestamp (default: current time)

Returns: Retrievability (0-1)

createEmptyCard(now?: Date): Card

Creates a new empty card ready for first review.

Parameters:

• now: Optional timestamp (default: current time)

Returns: New card in "New" state

getNextInterval(card: Card, rating: Rating, now?: Date): number

Gets the next interval for a given rating (preview mode).

Parameters:

• card: Card to evaluate
• rating: Rating to preview
• now: Current timestamp (default: current time)

Returns: Interval in days

Types

Rating

type Rating = 1 | 2 | 3 | 4; // Again | Hard | Good | Easy

const Rating = {
  Again: 1,
  Hard: 2,
  Good: 3,
  Easy: 4,
} as const;

State

type State = 0 | 1 | 2 | 3; // New | Learning | Review | Relearning

const State = {
  New: 0,
  Learning: 1,
  Review: 2,
  Relearning: 3,
} as const;

Card

interface Card {
  due: Date;              // Next scheduled review timestamp
  stability: number;      // Days to 90% retention
  difficulty: number;     // Complexity (1-10)
  elapsed_days: number;   // Days since last review
  scheduled_days: number; // Assigned interval
  reps: number;           // Total review count
  lapses: number;         // Failure count
  state: State;           // Current learning phase
  last_review: Date | null; // Previous review timestamp
}

FSRSParameters

interface FSRSParameters {
  request_retention: number;  // 0.7-0.97, default 0.9
  maximum_interval: number;   // Default 36500 days
  w: number[];                // 17 (v4.5) or 21 (v6) weights
  enable_fuzz: boolean;       // Random interval jitter
  enable_short_term?: boolean; // V6 same-day logic
}

Utility Functions

isLapse(grade: Rating | ContinuousRating): boolean

Checks if a grade represents a lapse (failure to recall). Useful for analytics, streak tracking, and UI badges.

Parameters:

• grade: Rating (discrete or continuous)

Returns: true if grade < 1.5 (Again)

Configuration

Default v4.5 Parameters

const fsrs = new FSRS({
  request_retention: 0.9,
  maximum_interval: 36500,
  enable_fuzz: true
});

Custom V6 Parameters (Advanced)

For users with optimized parameters from the FSRS optimizer:

const fsrs = new FSRS({
  request_retention: 0.9,
  w: [
    // Your 21 optimized parameters from FSRS optimizer
    0.4, 0.6, 2.4, 5.8, 4.93, 0.94, 0.86, 0.01,
    1.49, 0.14, 0.94, 2.18, 0.05, 0.34, 1.26,
    0.29, 2.61, 0.48, 1.05, 0.36, -0.54
  ]
});

Note: V6 parameters only provide benefit when optimized on your specific review data. Generic v6 defaults offer minimal improvement over v4.5.

Integration Examples

React + Zustand

import create from 'zustand';
import { FSRS, Card, Rating } from '@squeakyrobot/fsrs';

const fsrs = new FSRS();

interface FlashcardStore {
  cards: Map<string, Card>;
  reviewCard: (id: string, rating: Rating) => void;
}

const useFlashcards = create<FlashcardStore>((set) => ({
  cards: new Map(),
  reviewCard: (id, rating) => set((state) => {
    const card = state.cards.get(id);
    if (!card) return state;

    const result = fsrs.repeat(card, new Date());
    const updatedCard = result[rating].card;

    return {
      cards: new Map(state.cards).set(id, updatedCard)
    };
  })
}));

Next.js API Route

import { FSRS, Rating } from '@squeakyrobot/fsrs';
import { NextRequest, NextResponse } from 'next/server';

const fsrs = new FSRS();

export async function POST(req: NextRequest) {
  const { card, rating } = await req.json();

  const result = fsrs.repeat(card, new Date());
  const updated = result[rating as Rating];

  return NextResponse.json({
    card: updated.card,
    log: updated.log
  });
}

Performance Benchmarks

Based on the fsrs-benchmark dataset (1.7B reviews from 20K users):

Algorithm	Log Loss	RMSE	Improvement vs SM-2
SM-2 (Anki Default)	0.7317	0.4066	-
FSRS v4.5	0.3624	0.0764	~81%
FSRS v6	0.3460	0.0653	~84%

Key Insight: FSRS v4.5 captures 98% of v6's improvement with 19% fewer parameters, making it ideal for applications that don't have user-specific optimization.

Contributing

Contributions are welcome! See the CONTRIBUTING guide for development guidelines.

License

MIT License - see LICENSE for details.

References

• FSRS Official Documentation
• Technical Explanation of FSRS
• FSRS Benchmark Results
• Original TypeScript Implementation