---

name: elevenlabs-twilio-memory-bridge description: "FastAPI personalization webhook that adds persistent caller memory and dynamic context injection to ElevenLabs Conversational AI agents on Twilio. No audio proxying, file-based persistence, OpenClaw compatible." version: "1.2.0" author: britrik tags: ["elevenlabs", "twilio", "voice-agent", "telephony", "conversational-ai", "memory-injection", "fastapi"] metadata: openclaw: requires: env: - ELEVENLABS_API_KEY - ELEVENLABS_AGENT_ID - OPENCLAW_API_BASE_URL - PUBLIC_BASE_URL - ADMIN_API_KEY bins: - python3 - pip - uvicorn primaryEnv: ELEVENLABS_API_KEY envVars: - name: ELEVENLABS_API_KEY required: true description: Scoped ElevenLabs API key for agent access. - name: ELEVENLABS_AGENT_ID required: true description: ElevenLabs Conversational AI Agent ID. - name: OPENCLAW_API_BASE_URL required: true description: Base URL of your OpenClaw instance (HTTPS). - name: PUBLIC_BASE_URL required: true description: Publicly reachable URL of this bridge service. - name: ADMIN_API_KEY required: true description: Secret key for admin endpoint authentication (Bearer token). - name: WEBHOOK_SECRET required: true description: Shared secret for webhook HMAC verification. Must be set in production. - name: PHONE_HASH_SALT required: true description: Salt for HMAC-SHA256 phone number hashing. Generate with openssl rand -base64 32. - name: DATA_ENCRYPTION_KEY required: false description: Fernet key for encrypting memories and notes at rest. When unset, data is stored as plain JSON. - name: SOUL_TEMPLATE_PATH required: false description: Path to personality template file (default ./soul_template.md). - name: DATA_DIR required: false description: Directory for JSON persistence (default ./data). - name: ALLOWED_ORIGINS required: false description: Comma-separated CORS origins (leave unset to disable CORS). - name: HOST required: false description: Bind address for uvicorn (default 0.0.0.0). - name: PORT required: false description: Listen port (default 8000). - name: LOG_LEVEL required: false description: Python logging level (default INFO). emoji: "\U0001F4DE" homepage: https://github.com/britrik/elevenlabs-twilio-memory-bridge

elevenlabs-twilio-memory-bridge

Personalization webhook service for ElevenLabs + Twilio voice agents with persistent caller memory.

What It Does

When a call arrives on your Twilio number, ElevenLabs' native integration triggers this webhook. The bridge looks up the caller's history, loads long-term memory facts and daily context notes, combines them with a customizable soul/personality template, and returns everything as a system prompt override so your agent greets each caller with full context.

Architecture

• No audio proxying — ElevenLabs and Twilio handle media directly
• Webhook only — called once per inbound call to inject context
• File-based persistence — JSON files in ./data/, zero external dependencies
• OpenClaw compatible — works with any OpenAI-compatible LLM endpoint

Endpoints

Setup

1. Clone repo, pip install -r requirements.txt
2. Copy .env.example to .env, fill in secrets
3. Configure ElevenLabs agent with Custom LLM pointing to your OpenClaw instance
4. Enable system prompt + first message overrides in agent Security settings
5. Add webhook URL https://your-domain/webhook/personalize in ElevenLabs settings
6. Import Twilio number in ElevenLabs dashboard
7. Run: uvicorn app:app --host 0.0.0.0 --port 8000

Required Environment Variables

• ELEVENLABS_API_KEY — scoped ElevenLabs key
• ELEVENLABS_AGENT_ID — your agent ID
• OPENCLAW_API_BASE_URL — your OpenClaw instance URL
• PUBLIC_BASE_URL — publicly reachable URL of this service
• ADMIN_API_KEY — secret for admin endpoint auth
• PHONE_HASH_SALT — salt for phone number hashing. Generate with openssl rand -base64 32. Changing this invalidates all existing caller data.
• WEBHOOK_SECRET — shared secret for HMAC webhook verification. Must be set in production.

Optional Environment Variables

Security

• Deterministic phone hashing — Phone numbers are hashed with HMAC-SHA256 using PHONE_HASH_SALT as the key before storage. Changing the salt invalidates existing caller data.
• Encryption at rest — Memories and notes are encrypted with Fernet when DATA_ENCRYPTION_KEY is set. Data is stored as plain JSON when the key is not configured.
• Mandatory webhook verification — WEBHOOK_SECRET is required in production. Webhooks are rejected when the secret is not set (fail-closed).
• Rate limiting — Admin endpoints (/api/memory, /api/notes) are limited to 30 requests per minute. Webhook endpoints are limited to 60 per minute.
• File locking — JSON reads and writes use POSIX advisory locks (fcntl) to prevent race conditions under concurrent access.
• Restrictive permissions — Data directory (DATA_DIR) is created with 0o700 permissions. JSON files are written with 0o600.
• Secrets from env vars only — No hardcoded credentials in source.
• Safe for public GitHub repos — All sensitive config is loaded from environment variables.

Migration from v1.1.x to v1.2.x

v1.2.0 introduces breaking security changes:

1. PHONE_HASH_SALT is now required — Previously, phone numbers were hashed with unsalted SHA-256. Now HMAC-SHA256 with a salt is required. Existing caller data will be unreachable after upgrading because the hash keys change. You must accept data loss and start fresh (v1.2.0 uses a different hashing algorithm).

2. WEBHOOK_SECRET is now required — Previously optional for development. Now mandatory in all environments. Webhooks without a valid secret are rejected.

3. DATA_ENCRYPTION_KEY is optional but recommended — When set, all new memories and notes are encrypted at rest. Existing unencrypted data remains readable.

Migration checklist

1. Generate a salt: openssl rand -base64 32
2. Generate a webhook secret: openssl rand -hex 32
3. Generate an encryption key (optional): python -c "from cryptography.fernet import Fernet; print(Fernet.generate_key().decode())"
4. Add all to .env
5. Review CORS settings — ALLOWED_ORIGINS now defaults to disabled
6. Update ElevenLabs webhook settings with the new WEBHOOK_SECRET
7. Note: v1.2.0 requires fresh data — existing caller sessions will be recreated on first call