DeepTutor：终身个性化 AI 家教，真正记住你学到哪了，多 RAG 引擎+长期记忆+Skills 生态，Docker 自托管

by Mycelium Protocol

---

GitHub：HKUDS/DeepTutor
论文：arxiv.org/abs/2604.26962
官方文档：deeptutor.info
许可证：Apache 2.0
语言：Python 3.11+ + Next.js 16
Stars：36,949 · Forks：4,632
最新版本：v1.5.16（2026-08-22）
机构：香港大学数据科学实验室（HKUDS）

---

一、它和通用 AI 问答的本质区别

普通 AI 问答每次对话都从零开始——你问它线性代数，它不知道你上周刚学完矩阵乘法，卡在特征值上。

DeepTutor 的设计核心是长期记忆：三层持久化记忆结构让系统真正知道你学到哪一步了，不是靠上下文，而是跨会话的真实追踪。

这就是用户说的「它是真真切切知道你学到哪一步了」的技术底座。

---

二、三层记忆体系

层级	名称	内容
L1	轨迹层	每次会话的详细行为记录
L2	摘要层	从轨迹中提炼的知识状态
L3	综合层	跨时间的学习模型（你擅长什么、卡在哪里）

Memory Graph：每个记忆声明都能追溯到具体的证据来源，不是黑盒推断，而是可检查、可编辑的知识图谱。

---

三、七种学习模式，共用同一个 Agent 循环

DeepTutor 的独特架构：Chat、Quiz、Research、Visualize、Solve、Mastery Path、沉浸阅读这七种模式运行在同一个 Agent 引擎上。切换模式时，学习上下文完整保留——不是换了个工具，是换了个目标。

模式	用途
Chat	对话式问答，结合知识库检索
Quiz	AI 出题，自动批改，结果进入 Question Bank
Research / Deep Research	多步骤研究，跨文档综合
Visualize	概念可视化（Chart.js/SVG/Mermaid）
Solve	解题，带过程展示
Mastery Path	有掌握度门槛的结构化学习路径
沉浸阅读	文档在侧边展开，逐页引用，边读边问

---

四、多引擎知识库

不同文档类型、不同检索需求，接不同引擎：

引擎	特点
LlamaIndex	通用文档 RAG，支持多模态
PageIndex	按页检索，可推理，支持自托管
GraphRAG	知识图谱结构检索
LightRAG / LightRAG Server	轻量高速，支持远程服务
Obsidian Vault	直接链接本地笔记库
Tencent IMA	腾讯 IMA 库集成
MarginNote 4	读书笔记库（v1.5.16 新增）

文档解析引擎可插拔：LiteParse、Apache Tika（v1.5.15 新增）、PyMuPDF4LLM、MinerU。

---

五、Skills 生态

deeptutor skill install <skill-name>

技能来自 EduHub（ClawHub 社区），用户也可以自己写技能。v1.2.2 起支持用户创作技能，v1.4.4 起支持从社区安装。

内置工具之外还有：

• MCP 服务器：任意 MCP 工具接入
• CLI Apps：101 个内置命令行工具可供 Agent 调用
• 图片/视频/语音生成模型

---

六、Partners 系统

Partners 是带独立记忆和技能的持久化 AI 同伴，可以在任意对话轮次召唤：

Claude Code / Codex CLI / Gemini CLI / Kimi Code / opencode / MiMo ...

支持 15 个 IM 渠道实时串流，也可以把 Partners 的历史对话导入当前会话。

---

七、安装

方式一：pip（最快）

mkdir -p my-deeptutor && cd my-deeptutor
pip install -U deeptutor
deeptutor init     # 配置端口、LLM、Embedding
deeptutor start    # 启动后访问 http://127.0.0.1:3782

方式二：Docker

docker pull hkuds/deeptutor:latest
docker run -p 3782:3782 -p 8001:8001 \
  -v $(pwd)/data:/app/data \
  hkuds/deeptutor:latest

方式三：源码开发

git clone https://github.com/HKUDS/DeepTutor.git && cd DeepTutor
python3 -m venv .venv && source .venv/bin/activate
pip install -e .
cd web && npm ci --legacy-peer-deps && cd ..
deeptutor init && deeptutor start --dev

---

八、最近更新（一周内 4 个版本）

v1.5.16（2026-08-22，今日）：MarginNote 4 库集成；修复工具调用 ID、Embedding 和网关温度限制问题。

v1.5.15（2026-08-20）：PageIndex OSS 自托管版（带推理检索）；Question Bank 文件归档；第三方工具/能力插件；Apache Tika 文档解析。

v1.5.14（2026-08-19）：沉浸阅读（文档侧边展开，逐页引用）；从聊天直接配置 DeepTutor；Tencent IMA 库；Notebook 控制台。

v1.5.13（2026-08-17）：Book 流式编译 + 进度追踪 + 导出 Markdown；审批前费用预估；首页建议从记忆中生成。

---

九、成长轨迹

• 2025-12-29 首发
• 2026-02-06：10K stars，仅用 39 天
• 2026-04-19：20K stars，111 天
• 2026-08-22：36,949 stars

香港大学 HKUDS 实验室出品，有论文（arXiv 2604.26962）支撑，不是纯工程项目。

---

DeepTutor 的核心赌注是：真正的个性化学习需要跨会话的持久化记忆，而不是更大的上下文窗口。 三层记忆 + 可视化记忆图谱 + Mastery Path 的组合，让「AI 真的知道你学到哪了」从口号变成可验证的系统行为。

---

Mycelium Protocol — 追踪 AI 系统的底层演化

---

关于 Mycelium

菌丝协议。持续追踪 AI 工具、系统和实验的内容节点。

---

DeepTutor: Lifelong Personalized AI Tutoring — Multi-RAG Engines, Three-Layer Memory, Skills Ecosystem, Docker Self-Host

by Mycelium Protocol

---

GitHub: HKUDS/DeepTutor
Paper: arxiv.org/abs/2604.26962
Docs: deeptutor.info
License: Apache 2.0
Stack: Python 3.11+ + Next.js 16
Stars: 36,949 · Forks: 4,632
Latest: v1.5.16 (2026-08-22, today)
Institution: HKU Data Science Lab (HKUDS)

---

The Core Difference from Generic AI Q&A

Generic AI Q&A starts fresh every session — it doesn’t know you spent last week stuck on eigenvalues after learning matrix multiplication.

DeepTutor’s design is centered on long-term memory: a three-layer persistent memory structure that genuinely tracks where each learner is — not through context windows, but through real cross-session state.

This is the technical foundation behind the claim that “it actually knows where you are in your learning.”

---

Three-Layer Memory

Layer	Name	Content
L1	Trace	Detailed per-session behavior records
L2	Surface	Knowledge state distilled from traces
L3	Synthesis	Cross-time learning model (strengths, blockers)

Memory Graph: every memory claim traces back to specific evidence. Not a black box — inspectable and editable.

---

Seven Learning Modes, One Agent Loop

Chat, Quiz, Research, Visualize, Solve, Mastery Path, and Immersive Reading all run on the same agent engine. Switching modes preserves learning context — you’re changing the objective, not the tool.

Mode	Purpose
Chat	Conversational Q&A with RAG retrieval
Quiz	AI-generated questions, auto-graded, saved to Question Bank
Research / Deep Research	Multi-step synthesis across documents
Visualize	Concept visualization (Chart.js / SVG / Mermaid)
Solve	Step-by-step problem solving
Mastery Path	Structured learning with a hard mastery gate
Immersive Reading	Document open beside the thread, cited page by page

---

Multi-Engine Knowledge Base

Different document types and retrieval needs get different engines:

Engine	Strength
LlamaIndex	General-purpose RAG, multimodal
PageIndex	Page-level retrieval with reasoning; self-hostable
GraphRAG	Knowledge graph structure
LightRAG / LightRAG Server	Lightweight, fast, remote-capable
Obsidian Vault	Direct link to local note vault
Tencent IMA	IMA library integration
MarginNote 4	Reading annotation library (v1.5.16)

Pluggable document parsing: LiteParse, Apache Tika (v1.5.15), PyMuPDF4LLM, MinerU.

---

Skills Ecosystem

deeptutor skill install <skill-name>

Skills come from EduHub (ClawHub community), and users can author their own. Community install available since v1.4.4.

Beyond built-in tools:

• MCP servers: any MCP tool
• CLI Apps: 101 built-in command-line apps the agent can invoke
• Image / video / voice generation models

---

Partners

Partners are persistent AI companions with their own memory and skills, callable from any conversation turn:

Claude Code / Codex CLI / Gemini CLI / Kimi Code / opencode / MiMo ...

15 IM channels with live streaming; Partners’ conversation history can be imported into the current session.

---

Install

Option 1: pip (fastest)

mkdir -p my-deeptutor && cd my-deeptutor
pip install -U deeptutor
deeptutor init     # configure ports, LLM, embedding
deeptutor start    # open http://127.0.0.1:3782

Option 2: Docker

docker pull hkuds/deeptutor:latest
docker run -p 3782:3782 -p 8001:8001 \
  -v $(pwd)/data:/app/data \
  hkuds/deeptutor:latest

Option 3: source

git clone https://github.com/HKUDS/DeepTutor.git && cd DeepTutor
python3 -m venv .venv && source .venv/bin/activate
pip install -e .
cd web && npm ci --legacy-peer-deps && cd ..
deeptutor init && deeptutor start --dev

---

Recent Velocity (4 releases in 7 days)

v1.5.16 (2026-08-22, today): MarginNote 4 library; tool-call ID, embedding, and gateway temperature fixes.

v1.5.15 (2026-08-20): Self-hosted PageIndex OSS with reasoning retrieval; Question Bank filing; third-party capability plugins; Apache Tika parsing.

v1.5.14 (2026-08-19): Immersive Reading (document beside thread, page-by-page citations); chat-driven self-configuration; Tencent IMA library; notebook console.

v1.5.13 (2026-08-17): Books stream while compiling, progress tracking, Markdown export; cost estimate before approval; memory-driven home suggestions.

---

Growth

• 2025-12-29: first release
• 2026-02-06: 10K stars in 39 days
• 2026-04-19: 20K stars in 111 days
• 2026-08-22: 36,949 stars

From HKU’s Data Science Lab with a peer-reviewed paper (arXiv 2604.26962) — research-backed, not just engineering.

---

DeepTutor’s core bet: real personalized learning needs persistent cross-session memory, not a larger context window. Three-layer memory + inspectable Memory Graph + Mastery Path gate turns “the AI knows where you are” from a tagline into verifiable system behavior.

---

Mycelium Protocol — tracking the deep evolution of AI systems

© 2026 Mycelium Protocol. All rights reserved.