selfOS

Compile your notes, conversations, and bookmarks into a structured, queryable knowledge graph.

A Claude Code skill suite that turns scattered personal data into a three-layer markdown wiki with cross-references and citations.

[Screenshot: Knowledge graph visualization showing interconnected concepts, entities, and sources]

Requires Claude Code (npm install -g @anthropic-ai/claude-code)

English | 中文

---

What is selfOS?

Most personal knowledge tools use RAG -- they re-read raw files on every query, burning tokens and losing structure. selfOS takes a different approach: compilation.

Each source is processed once into a three-layer wiki:

• Source -- one summary per ingested document
• Concept -- abstract ideas, methods, patterns (cross-referenced)
• Entity -- people, organizations, tools, papers

After compilation, querying costs zero LLM tokens. Just read 2-3 markdown files.

Key Features

• Knowledge Compilation -- Three-layer architecture: source, concept, entity. Process once, query forever. Zero token cost after ingestion.
• Context Recovery -- AI interviews YOU to fill gaps. /interview surfaces open questions and thin pages. /bookmark-chat recovers forgotten context from old bookmarks and terse notes.
• Quick Thought Capture -- /thought lets you jot down an idea and optionally run a follow-up interview in one command.
• Wiki Digest -- /digest generates a daily activity summary and recommends questions to deepen your knowledge base.
• Auto-Capture -- Stop hook detects valuable personal context in Claude Code conversations and silently saves to wiki. No manual tagging needed.
• Multi-Source Ingest -- Notion notes, Claude/Gemini conversations, Twitter bookmarks, PDFs, web pages. One command: /wiki ingest.

Command Reference

Wiki Operations

Command	Description
/wiki init	Scaffold directory structure, create CLAUDE.md, initialize git
/wiki ingest <url-or-path>	Fetch or copy a source, compile into source + concept + entity pages
/wiki query "<question>"	Search index, read relevant pages, synthesize answer with citations
/wiki lint	Health check: orphan pages, broken links, missing frontmatter, uncited claims
/wiki compile	Batch-ingest all unprocessed files in raw/
/wiki status	Show page counts, word counts, recent log entries
/wiki sync	Incremental sync from Notion to wiki

Context Recovery

Command	Description
/interview	Wiki asks YOU questions -- targets open questions, thin pages, timeline gaps
/bookmark-chat	Pick a random Twitter bookmark or terse thought, recover the context behind it
/complete	Same as bookmark-chat but draws only from Notion Thoughts lacking context

Capture and Review

Command	Description
/thought <text>	Jot down a quick thought, optionally triggers a follow-up interview
/digest	Daily activity summary + recommended deepening questions
/digest week	Weekly digest across all activity
/digest question	Generate targeted questions from current wiki state

Quick Start

Explore the Demo

git clone https://github.com/freemty/selfOS
cd selfOS

# Open in Claude Code
claude

# Query the included demo wiki
/wiki query "What is research taste?"

Start Fresh (Your Own Wiki)

git clone https://github.com/freemty/selfOS
cd selfOS
rm -rf wiki/sources/ wiki/concepts/ wiki/entities/ wiki/synthesis/  # clear demo content

# Open in Claude Code
claude

# Initialize and ingest your first source
/wiki init
/wiki ingest <your-first-source>

Your first source will generate pages in wiki/sources/, wiki/concepts/, and wiki/entities/.

Documentation

• Graph Viewer -- Interactive knowledge graph visualization
• Importing Conversations -- Claude, Gemini, Notion data import
• Obsidian Setup -- View your wiki in Obsidian

How It Works

raw/ (your data)
  |
  v
/wiki ingest --> wiki/sources/   (one summary per source)
                    |
                    v
               wiki/concepts/  (abstract ideas, cross-referenced)
               wiki/entities/  (people, orgs, tools)
                    |
                    v
               graph.json --> Interactive Visualization
                    |
                    v
               /wiki query --> Cited answers from compiled knowledge

Architecture

selfOS follows a compilation philosophy, not retrieval:

1. Source Layer (wiki/sources/) -- Each raw document is summarized once with metadata, key takeaways, and citations. The raw source is never re-read.

2. Concept Layer (wiki/concepts/) -- Abstract ideas, methods, and patterns are extracted and cross-referenced using [[wiki-links]]. Multiple sources can feed into a single concept page.

3. Entity Layer (wiki/entities/) -- People, organizations, tools, and papers get their own pages with structured frontmatter and bidirectional links.

All pages use YAML frontmatter for structured metadata, [[wiki-links]] for cross-references, and (source: [[sources/slug]]) for citations. The result is a static knowledge graph that can be browsed in Obsidian, queried by Claude Code, or visualized with the included Flask viewer (viewer/app.py).

Demo Data

The repository includes 9 fictional wiki pages built around "Alex Chen," a PhD student navigating research directions:

• 3 Sources: a Claude conversation about choosing systems over theory, a Gemini deep-dive on Flash Attention, a Notion note on research taste
• 3 Concepts: research taste, learning by building, career pivots
• 3 Entities: Tri Dao, a study group, an academic advisor

Explore them to understand the page format before building your own wiki:

# Browse demo pages in the viewer
python viewer/app.py    # opens localhost:5001

# Or read them directly
ls wiki/sources/ wiki/concepts/ wiki/entities/

Built With

Claude Code skill system -- HTML5 Canvas -- Flask

License

MIT

---

selfOS -- 中文说明

将笔记、对话、书签编译成结构化的、可查询的知识图谱。

一套 Claude Code skill，把零散的个人数据编译成三层 markdown wiki，带交叉引用和来源标注。

什么是 selfOS?

大部分个人知识工具使用 RAG -- 每次查询都重新读取原始文件，消耗大量 token，丢失结构。selfOS 走的是另一条路：编译。

每个来源只处理一次，生成三层 wiki：

• Source -- 每个文档一份摘要
• Concept -- 抽象概念、方法论、模式（交叉引用）
• Entity -- 人物、组织、工具、论文

编译完成后，查询零 token 消耗。只需读 2-3 个 markdown 文件。

核心功能

• 知识编译 -- 三层架构：source、concept、entity。处理一次，永久查询。
• Context Recovery -- AI 主动向你提问来填补空白。/interview 发现开放问题和薄弱页面。/bookmark-chat 从旧书签和简短笔记中恢复遗忘的 context。
• 快速记录 -- /thought 随手记下想法，可选触发一轮跟进采访。
• Wiki Digest -- /digest 生成每日活动摘要，推荐深化知识库的问题。
• 自动捕获 -- Stop hook 自动检测对话中有价值的个人 context，静默保存到 wiki。
• 多源导入 -- Notion 笔记、Claude/Gemini 对话、Twitter 书签、PDF、网页。一条命令：/wiki ingest。

命令速查

Wiki 操作

命令	说明
/wiki init	创建目录结构、CLAUDE.md、初始化 git
/wiki ingest <url-or-path>	获取来源，编译为 source + concept + entity 页面
/wiki query "问题"	搜索索引，读取相关页面，生成带引用的回答
/wiki lint	健康检查：孤页、断链、缺失 frontmatter、无引用声明
/wiki compile	批量编译 raw/ 中未处理的文件
/wiki status	显示页面数、字数、最近日志
/wiki sync	从 Notion 增量同步到 wiki

Context Recovery

命令	说明
/interview	Wiki 主动向你提问 -- 针对开放问题、薄弱页面、时间线空白
/bookmark-chat	随机抽一条书签或简短 thought，还原背后的 context
/complete	只从缺乏 context 的 Notion Thoughts 中抽取

捕获与回顾

命令	说明
/thought <text>	快速记录想法，可选触发跟进采访
/digest	每日活动总结 + 推荐深化问题
/digest week	周度汇总
/digest question	从当前 wiki 状态生成针对性问题

快速开始

前置要求：Claude Code（npm install -g @anthropic-ai/claude-code）

体验 Demo

git clone https://github.com/freemty/selfOS
cd selfOS

# 在 Claude Code 中打开
claude

# 查询内置的 demo wiki
/wiki query "什么是 research taste？"

从零开始（构建自己的 wiki）

git clone https://github.com/freemty/selfOS
cd selfOS
rm -rf wiki/sources/ wiki/concepts/ wiki/entities/ wiki/synthesis/  # 清除 demo 内容

# 在 Claude Code 中打开
claude

# 初始化并导入第一个来源
/wiki init
/wiki ingest <你的第一个来源>

第一个来源会自动在 wiki/sources/、wiki/concepts/、wiki/entities/ 下生成页面。

文档

• Graph Viewer -- 交互式知识图谱可视化
• 导入对话 -- Claude、Gemini、Notion 数据导入
• Obsidian 配置 -- 在 Obsidian 中浏览 wiki

演示数据

仓库包含 9 个虚构的 wiki 页面，围绕 "Alex Chen"（一个探索研究方向的博士生）展开：

• 3 个 Source：关于选择系统方向的 Claude 对话、Flash Attention 深入分析、关于 research taste 的 Notion 笔记
• 3 个 Concept：research taste、learning by building、career pivots
• 3 个 Entity：Tri Dao、study group、academic advisor

# 用可视化工具浏览
python viewer/app.py    # localhost:5001

# 或直接查看文件
ls wiki/sources/ wiki/concepts/ wiki/entities/

许可证

MIT