MemVoice Voice Agent Starter

Realtime voice assistant starter that pairs a Pipecat backend (Python) with a Next.js UI deployed on Vercel. Models are switchable per session (OpenAI, Gemini, Claude) and long‑term memory is handled by Mem0 so context survives across model hops.

Quick start

Web (Next.js)

1. cd web
2. npm install
3. Set NEXT_PUBLIC_BOT_URL=http://localhost:7860 in web/.env.local
4. npm run dev and open http://localhost:3000

Bot (Pipecat + Mem0)

1. python -m venv .venv && .venv/Scripts/activate (or source .venv/bin/activate)
2. pip install -r bot/requirements.txt
3. Export the required API keys (see docs/CONFIG.md). If self-hosting Mem0, set MEM0_BASE_URL=https://mem0.yourdomain.com.
4. Run the server: python bot/main.py (exposes WebRTC endpoints for voice and /chat for text on port 7860)

Once both are up, connect from the UI, choose a provider/model, and start talking. Memory is keyed by your userId (stored locally in the browser) so you can swap models without losing context.

Repository layout

• /web — Next.js App Router UI with Pipecat client, provider/model selector, transcript, and connection controls.
• /bot — Python backend with Pipecat pipeline, Mem0 integration, model router, and runner entry point.
• /docs — Architecture, config, memory, and deployment notes.

Scripts

• npm --prefix web run dev — start the UI locally
• npm --prefix web run lint — lint the UI
• python bot/main.py -t webrtc — run Pipecat runner using Small WebRTC transport
• pytest bot/tests — quick Python test for Mem0 wrapper

More detail in docs/DEPLOYMENT.md and docs/ARCHITECTURE.md.