Implementing FSRS Spaced Repetition in Flutter + Supabase
The Problem
My app "AI University" lets users learn about 66+ AI providers through quizzes and news. The problem: how do you make learning stick?
The answer: FSRS (Free Spaced Repetition Scheduler) — the algorithm powering next-gen flashcard systems like the latest Anki. Here's how I integrated it into a Flutter + Supabase stack.
What is FSRS?
FSRS models memory using two parameters:
- Stability (S): How well the memory is consolidated. Higher S = longer interval before next review.
- Difficulty (D): Subjective difficulty of the card.
After each review, the algorithm calculates a new interval based on a target retrievability (e.g., 90% recall probability).
Database Schema
CREATE TABLE ai_university_fsrs_cards (
id uuid PRIMARY KEY DEFAULT gen_random_uuid(),
user_id uuid REFERENCES auth.users NOT NULL,
provider text NOT NULL, -- 'google', 'openai', etc.
question_id text NOT NULL,
due_date timestamptz NOT NULL DEFAULT now(),
stability float NOT NULL DEFAULT 1.0,
difficulty float NOT NULL DEFAULT 0.3,
state text NOT NULL DEFAULT 'new', -- new/learning/review/relearning
reps int NOT NULL DEFAULT 0,
lapses int NOT NULL DEFAULT 0,
last_review timestamptz,
UNIQUE(user_id, provider, question_id)
);
-- Each user only sees their own cards
ALTER TABLE ai_university_fsrs_cards ENABLE ROW LEVEL SECURITY;
CREATE POLICY "users_own_fsrs_cards" ON ai_university_fsrs_cards
FOR ALL USING (user_id = auth.uid());
Edge Function: Two Actions
I added two actions to the existing ai-hub Edge Function:
quiz.fsrs_next — Get today's due cards
case "quiz.fsrs_next": {
const { provider, limit = 10 } = body;
const { data: cards } = await supabase
.from("ai_university_fsrs_cards")
.select("*")
.eq("user_id", user.id)
.eq("provider", provider)
.lte("due_date", new Date().toISOString())
.order("due_date")
.limit(limit);
return new Response(JSON.stringify({ success: true, cards }));
}
quiz.fsrs_grade — Record answer and schedule next review
case "quiz.fsrs_grade": {
const { provider, question_id, grade } = body; // grade 1-4
const newStability = calculateFsrsStability(currentCard, grade);
const intervalDays = Math.round(newStability * retrievabilityFactor);
const nextDue = new Date(Date.now() + intervalDays * 86400000);
await supabase.from("ai_university_fsrs_cards").upsert({
user_id: user.id, provider, question_id,
stability: newStability,
due_date: nextDue.toISOString(),
state: grade === 1 ? "relearning" : "review",
reps: currentCard.reps + 1,
last_review: new Date().toISOString()
});
return new Response(JSON.stringify({ success: true, next_due: nextDue }));
}
Flutter Service Layer
class AiFsrsService {
final _supabase = Supabase.instance.client;
// grade: 1=Again, 2=Hard, 3=Good, 4=Easy
Future<({DateTime nextDue, double stability})> gradeCard({
required String provider,
required String questionId,
required int grade,
}) async {
final response = await _supabase.functions.invoke('ai-hub', body: {
'action': 'quiz.fsrs_grade',
'provider': provider,
'question_id': questionId,
'grade': grade,
});
final nextDue = DateTime.parse(response.data?['next_due'] ?? '');
final stability = (response.data?['stability'] as num?)?.toDouble() ?? 1.0;
return (nextDue: nextDue, stability: stability);
}
static String nextDueLabel(DateTime nextDue) {
final diff = nextDue.difference(DateTime.now()).inDays;
if (diff <= 0) return 'Today';
if (diff == 1) return 'Tomorrow';
return 'In $diff days';
}
}
Home Card: Due Card Counter
After grading cards, I display a review reminder on the home screen:
// Query due cards directly from Supabase (RLS handles auth)
final dueCardRow = await _supabase
.from('ai_university_fsrs_cards')
.select('id')
.eq('user_id', user.id)
.lte('due_date', DateTime.now().toIso8601String())
.count(CountOption.exact);
setState(() => _dueCardCount = dueCardRow.count);
When _dueCardCount > 0, a "Review X cards" badge and button appear on the home card — a subtle but effective re-engagement hook.
Bug Caught Along the Way
When writing admin RLS policies on user_profiles, I initially used:
-- WRONG: 'id' is not the auth-mapped column
EXISTS (SELECT 1 FROM user_profiles WHERE id = auth.uid() AND is_admin = true)
-- RIGHT: user_profiles uses 'user_id' to reference auth.users
EXISTS (SELECT 1 FROM user_profiles up WHERE up.user_id = auth.uid() AND up.is_admin = true)
This kind of mistake silently passes schema validation but blocks admin access entirely.
What's Next
- Voice + FSRS: Answer cards by speaking — Web Speech API grades the answer and FSRS schedules the next review.
-
Adaptive difficulty: Use the
learner_profilestable to route hard cards through a more capable LLM. - Cross-provider review sessions: A single session that surfaces due cards across multiple AI providers.
Building in public: https://my-web-app-b67f4.web.app/#/gemini-university
