Agent-native tutoring with memory layers and a CLI

What it does DeepTutor is an agent-native tutoring platform that wraps large language models, LlamaIndex-based RAG, and persistent memory into a single system. It offers a web interface and a CLI for interacting with autonomous TutorBot agents that can conduct deep research, co-author documents, generate interactive visualizations, and compile “living books” from knowledge bases. The project is built in Python with a Next.js frontend and supports both solo use and multi-user shared deployments.

The interesting bit Instead of bolting a chat UI onto a generic LLM, DeepTutor treats tutoring as a multi-agent workflow built around a three-layer memory workbench (L1/L2/L3) and a plugin-style Tools + Capabilities engine. The system versions its knowledge-base indexes, persists state across restarts, and routes reasoning tasks through an agentic engine rather than a single prompt chain.

Key highlights

• Three-layer memory workbench (L1/L2/L3) with a unified Space context for chat history, skills, and memory
• Persistent autonomous TutorBots with sandboxed tool use and an HTTP/SSE API
• LlamaIndex-only RAG pipeline with versioned knowledge-base indexes and re-index workflows
• Multi-user deployments with isolated workspaces, auth routes, and scoped runtime access
• Broad model support including OpenAI, Anthropic, Gemini, NVIDIA NIM, vLLM, LM Studio, and llama.cpp
• Agentic capabilities: Deep Research, Solve, Question, Visualize/Animator, and Co-Writer

Verdict Worth exploring if you need a tutoring stack with agentic memory and RAG that deploys locally or via Docker. Avoid it if you want a lightweight chat UI without the backend complexity.