Mnemo AI

A local agentic AI assistant with MCP (Model Context Protocol) integration, RAG capabilities, and intelligent conversation management. Built on LangGraph with LangChain for multi-provider LLM support (Ollama, MLX, Amazon Bedrock, Bedrock Mantle, OpenAI, Anthropic, Amazon SageMaker AI, LiteLLM).

assistant-demo.mp4

📖 Documentation

Full documentation is available at https://brunopistone.github.io/mnemoai/

• Getting Started
• Usage & commands
• Configuration
• Orchestration & sub-agents
• Memory & learning
• Everyday tools
• Architecture
• Development

🚀 Quick Start

pip install mnemoai-assistant   # or: uv tool install mnemoai-assistant
mnemoai                          # verbose (shows thinking); --no-verbose to hide

On first run, if no config is found, an interactive configurator launches and walks you through picking a provider, model, and feature toggles — then writes ~/.mnemoai/config/config.yaml.

→ See the Getting Started guide for full setup.

✨ Key Features

• 🤖 Multi-Model Support: Ollama (local), MLX (local, Apple Silicon), Amazon Bedrock, Bedrock Mantle, OpenAI, Anthropic (Claude), Amazon SageMaker AI (chat only — no tool calling), LiteLLM (100+ providers)
• 🔧 MCP Tool System: Extensible tool architecture via Model Context Protocol
• 📚 RAG: Automatic document indexing and semantic (hybrid) search
• 🧠 User Profile Learning: Personalized responses learned from interactions
• 🧩 Episodic Memory: Learns from successful task completions and retrieves similar solutions
• 📖 ACE Playbook: Learns strategies from successes AND failures (Agentic Context Engineering)
• 🔍 Web Search & 🌐 Crawler: Brave Search API + web page extraction with RAG ingestion
• 🖼️ Vision Support: Image analysis with vision models
• 📁 File Operations & ✏️ Precise Editing: Read/write/edit text, CSV, JSON, PDF, DOCX
• 🔎 Fast Search: Glob + ripgrep content search (10-100x faster)
• 📋 Todo Tracking, 📝 Plan Mode & 🔄 Background Tasks: Multi-step task management
• ⚡ Bash Execution & 🛡️ Git Safety: Shell commands with smart error handling and guardrails
• 🪝 Tool Hooks: Your own commands run before/after any tool call — format writes, deny paths, auto-approve safe commands (docs)
• ⌨️ Your Own Slash Commands: A markdown file per prompt you retype — commands/review.md becomes /review <path> (docs)
• 📎 @-File Mentions: Type @ to complete a path anywhere in your prompt; the file is read and sent with the question (docs)
• ⟲ Take Back a Prompt: /rewind drops your last prompt and everything the turn produced — the conversation only, files on disk untouched (docs)
• 🗂️ Workspace Reports: /files lists what the session read, changed or attached; /diff shows uncommitted changes with this session's edits marked; /why <path> names the prompt behind each change to a file, across sessions; /copy puts the last answer on the clipboard without the terminal's wrapping (docs)

📄 License

Licensed under the MIT License — see the LICENSE file for details.

🤝 Contributing

This is a personal development project. Feel free to fork and adapt it to your needs; attribution to the original repository is appreciated but not required.