DeepTutor: The Agent-Native AI Tutor That Actually Remembers What You Don't Know

By Prahlad Menon 5 min read

Every AI study tool today has the same problem: it forgets you.

Ask ChatGPT to help with organic chemistry on Monday. Come back Wednesday for the next chapter. It has no idea you spent 40 minutes struggling with nucleophilic substitution two days ago. You’re starting from scratch. Every. Single. Time.

DeepTutor from HKU’s Data Intelligence Lab fixes this with an architecture that treats memory as a first-class citizen — not an afterthought bolted onto a chat wrapper.

What Makes This Different

DeepTutor isn’t a chatbot with a fancy prompt. It’s a four-layer agent system built from the ground up around persistent learning:

  1. User Interface Layer — web UI, CLI, and SDK for structured bidirectional interaction
  2. Specialized Agent Modules — problem solving (with a dual-loop architecture), deep research, guided learning, idea generation
  3. Tool Integration Layer — RAG pipelines, web search (Brave, Tavily, DuckDuckGo, SearXNG, and more), code execution
  4. Knowledge Graph Memory Foundation — this is the key layer. A persistent graph that tracks your learning progress across sessions, features, and even different tutoring agents

The v1.0.0 release (April 4, 2026) was a ground-up rewrite — roughly 200,000 lines of code with a two-layer plugin model. It’s already at 15,800+ stars and 2,000+ forks on GitHub. Apache 2.0 licensed.

The Unified Chat Workspace

Most AI education tools make you switch between different apps for different tasks. DeepTutor runs five modes in a single thread, all sharing context:

  • Chat — standard conversation with your tutor
  • Deep Solve — step-by-step problem solving with a dual-loop approach (attempt, verify, retry)
  • Quiz Generation — creates assessments based on what you’ve been studying
  • Deep Research — multi-source investigation that pulls from your knowledge base and the web
  • Math Animator — visual mathematical explanations

The important part: these aren’t separate tools. They share the same conversation thread and learner context. When you generate a quiz after a Deep Solve session, the quiz knows exactly which concepts you struggled with.

Personal TutorBots

This is where things get interesting. DeepTutor lets you create autonomous tutoring agents — each with their own memory, personality, and skill sets. Want a strict Socratic tutor for philosophy and a patient, visual-first tutor for calculus? Build both.

These TutorBots are powered by nanobot (also from the HKUDS lab, 38,000+ stars), and they can operate across channels: Telegram, Discord, Slack, WeChat, WhatsApp, Feishu, DingTalk. Your calculus tutor can quiz you on Telegram during your commute, and the session feeds back into your unified learner profile.

The Memory Architecture

This is the core innovation. DeepTutor builds a living learner profile:

  • What you’ve studied — every document, session, and interaction
  • How you learn — which explanations clicked, which didn’t
  • Where you’re heading — your learning goals and trajectory
  • What you struggle with — persistent tracking of weak areas

This profile is shared across all features and all TutorBots. When you upload a new PDF to the Knowledge Hub, the system already knows your baseline. When a TutorBot generates practice problems, it targets your actual gaps — not generic weak spots.

The Knowledge Graph Memory Foundation means relationships between concepts are preserved, not just flat embeddings. If you understand derivatives but struggle with integration by parts, the system knows these are related and can bridge the gap intelligently.

Knowledge Hub and AI Co-Writer

Upload PDFs, Markdown files, or plain text, and DeepTutor builds RAG-ready knowledge bases organized in color-coded notebooks. This isn’t just document storage — it’s structured, searchable, and integrated with every other feature.

The AI Co-Writer is a Markdown editor where AI is a first-class collaborator. It draws from your knowledge base and web sources simultaneously. Think of it as writing with a research assistant who’s already read everything you’ve uploaded.

Guided Learning

Feed DeepTutor your course materials and it transforms them into structured, multi-step visual learning journeys. Not just a summary — an actual pedagogical path with checkpoints, branching based on comprehension, and visual aids.

Agent-Native by Design

DeepTutor ships with a CLI where every capability is one command away. More importantly, it includes a SKILL.md file — which means other AI agents can programmatically operate DeepTutor. If you’re building an AI workflow that needs tutoring capabilities, DeepTutor is designed to be a tool, not just an app.

It supports 30+ LLM providers out of the box: OpenAI, Anthropic, Azure, Gemini, Ollama, vLLM, DeepSeek, and many more. Docker deployment is available for self-hosting.

Why This Matters for Medical Education

The persistent memory architecture has direct implications for medical education, CME, and clinical training. Medical learning is inherently sequential and cumulative — you can’t understand pharmacokinetics without understanding physiology. Current AI tools treat every session as isolated, which is the opposite of how medical knowledge builds.

A system that tracks a medical student’s learning trajectory — knowing they’ve mastered cardiac physiology but are shaky on renal — and builds the next study session accordingly? That’s genuinely useful. Combine that with the Knowledge Hub (upload your Pathoma notes, First Aid, and lecture slides into a unified RAG pipeline) and the Quiz Generation mode targeting actual weak areas, and you have something that could meaningfully improve board prep and clinical training.

The Bigger Picture

The AI education space is crowded with chat wrappers. Most “AI tutors” are a system prompt and a text box. DeepTutor is architecturally different:

  • Memory is structural, not conversational — a knowledge graph, not chat history
  • Personalization is earned, not declared — built from actual interaction patterns, not user preferences
  • The system is composable — CLI, SDK, multi-channel, agent-operable
  • It’s truly open source — Apache 2.0, self-hostable, no vendor lock-in

The gap between “AI that answers questions” and “AI that teaches you” is exactly this: memory, structure, and pedagogical awareness. DeepTutor is the most serious open-source attempt at closing that gap.

Links: