HomeBound is an AI-powered home search advisor/agent built for the WeHack 2026 CBRE Intelligent Memory Systems Challenge. It learns your preferences through conversation, remembers what matters across sessions, and surfaces real listings ranked to your profile using PyTorch embeddings and Claude.

Demo: https://youtu.be/VWFIbqfUpw8

• Frontend: React, JavaScript, CSS/Tailwind
• Backend: Python, Flask, SQLite, PyTorch, sentence-transformers
• AI: Anthropic Claude API (memory extraction + conversational agent)
• Listings: RapidAPI Realty in US (real MLS data)

1. run the following commands on bash to set up a virtual environment:

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

2. then create an .env file inside the backend folder and add the following keys

ANTHROPIC_API_KEY=your_anthropic_key_here
RAPIDAPI_KEY=your_rapidapi_key_here

3. Run the backend

cd backend
python app.py

Backend runs at: http://localhost:8080

1. Create the React app, install dependencies and run:

npm create vite@latest frontend -- --template react
cd frontend
npm install
npm run dev

Frontend runs at: http://localhost:5173

The agent builds a persistent memory profile for each user across three tiers:

• Explicit Memory — facts stated directly: budget, bedrooms, bathrooms, cities, property type, HOA preference
• Inferred Memory — priorities extracted from conversation: good schools, quiet neighborhood, big backyard, pool
• Evolving Memory — a full changelog of every update with old → new values and the reason for the change

All memory persists across sessions via SQLite. Users can update preferences by chatting naturally, and can selectively forget specific information on request.

1. RapidAPI returns up to 40 listings filtered by budget, beds, baths, year built, property type, and HOA preference
2. PyTorch sentence-transformers encode both the user memory profile and each property into vectors
3. Cosine similarity scores each property against the user profile
4. Soft penalties applied for constraint mismatches (wrong type, over budget, wrong bed/bath count)
5. Heavy penalty for dismissed properties, boost for liked properties
6. Top 9 results returned with match scores