NotForgotten 🧠

AI-powered spaced repetition learning system — combines semantic search with the Ebbinghaus forgetting curve to surface what you're about to forget, before you forget it.

---

Architecture

PDF → chunk → embed → ChromaDB (semantic store)
                           ↓
               HybridRetriever (α·semantic + (1-α)·urgency)
                           ↓
               LangGraph Agent (retrieve → generate → assess → update)
                           ↓
               SQLite (SM-2 memory) + networkx (concept graph)
                           ↓
               FastAPI (/session/* + /dashboard)

Quick Start

1. Install dependencies

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

2. Configure environment

cp .env.example .env
# Edit .env with your LLM_PROVIDER and API key

3. Start the API

uvicorn api.main:app --reload --port 8080

4. Ingest your own PDF

# Upload via curl or use the Frontend UI
curl -X POST http://localhost:8080/ingest -F "file=@your_document.pdf"

5. Begin a session

# Ask a question
curl -X POST http://localhost:8080/session/query \
  -H "Content-Type: application/json" \
  -d '{"query": "Summarize the key points", "session_id": "abc123"}'

# Rate your understanding (1-5) to update memory
curl -X POST http://localhost:8080/session/rate \
  -H "Content-Type: application/json" \
  -d '{"session_id": "abc123", "concept_id": "your_concept_id", "rating": 4}'

API Endpoints

Method	Endpoint	Description
POST	/session/start	Returns today's highest-urgency concepts
POST	/session/query	Runs full LangGraph pipeline, returns answer + sources
POST	/session/rate	Accepts self-rating, triggers SM-2 memory update
GET	/dashboard	Full memory state JSON for frontend

Key Concepts

• SM-2 Algorithm — Spaced repetition scheduler. Low ratings → shorter intervals. High ratings → longer intervals.
• Ebbinghaus Retention — R = e^(−t/S) where t = days since review, S = stability. Low retention = high urgency.
• Hybrid Retrieval — score = α × semantic_similarity + (1−α) × urgency. Chunks from forgotten concepts get boosted automatically.
• Re-Teach Mode — If you rate understanding ≤ 2, the agent loops back and retrieves additional context.