Self-hosted, event-driven Agentic AI assistant with triple-layer memory, dynamic persona, multi-channel support, and MCP ecosystem integration.
LimeBot is a local-first Agentic AI assistant framework for personal and community use. Built on an event-driven architecture with an asyncio.Queue message bus connecting four channels: Discord, WhatsApp, Telegram, and Web Dashboard.
Core cognitive capabilities are based on a triple-layer memory system: episodic memory (daily Markdown logs), semantic memory (LanceDB vector database), and Auto-RAG (semantic-first + grep fallback automatic retrieval), paired with a reflection engine that distills key facts into permanent memory every 4 hours. The dynamic persona system is defined through SOUL.md, IDENTITY.md, and MOOD.md files, supporting per-user relationship profiles and platform-specific tone differentiation.
Tool integration includes Playwright web browsing (navigation/search/forms/image download), sandboxed filesystem operations, secure shell execution (injection filtering + timeout termination), cron scheduled tasks, isolated sub-agent delegation, and a full MCP client supporting arbitrary external tool server extensions. The skill system uses SKILL.md + script units, with 6 built-in core skills and support for the ClawHub community registry and GitHub URL installation.
LLM access is unified through LiteLLM, supporting Gemini (default), OpenAI, Anthropic, xAI, DeepSeek, Qwen, and NVIDIA, with real-time model switching and AI Gateway proxy routing. The frontend uses React + Vite with WebSocket real-time streaming for tokens, tool execution cards, and confirmation prompts. Supports one-click startup, manual development, and Docker Compose deployment. MIT licensed.
System Requirements: Windows 10+ / macOS 11+ / Linux (Ubuntu 20.04+); Node.js 18+, Python 3.11–3.14; minimum 4GB RAM, ~2GB disk space.
Quick Start:
git clone https://github.com/LemonMantis5571/LimeBot.git
cd LimeBot
npm run lime-bot start
Docker Deployment:
docker-compose up --build
Unconfirmed: ClawHub registry URL, OpenClaw/Nanobot inspiration source links, whether an independent website exists, specific embedding model used for semantic memory.