LLM Knowledge OS

A Knowledge Operating System for AI-assisted personal knowledge management.

Built on top of Andrej Karpathy's LLM Wiki idea.

The Problem

Most personal knowledge systems stop at storing information.

Modern LLMs retrieve notes, summarize documents and answer questions, but they rarely build persistent understanding.

LLM Wiki introduced a compelling idea: let the LLM maintain an evolving wiki instead of repeatedly rediscovering knowledge.

This project extends that idea.

What is LLM Knowledge OS?

LLM Knowledge OS does not replace LLM Wiki. It extends it.

Raw Sources
    ↓
Wiki
    ↓
Knowledge Operating System
    ↓
AI Agents

Philosophy

Notes are documents. Knowledge Objects are software.

Quick Start

1. Clone the repository.

git clone https://github.com/giuseppecapasso/llm-knowledge-os

2. Open your empty Obsidian vault.

The .kos/ directory is hidden from Obsidian by default (dotfolders are invisible). This is intentional — the framework stays out of your way.

3. Open your AI coding agent (Claude Code, Codex, Cursor, Gemini CLI, OpenCode, etc.)

4. Run the bootstrap prompt.

Copy the contents of .kos/PROMPT.md and paste it into your agent.

5. The agent will:

• Read the framework from .kos/
• Create your vault structure at root: knowledge/, raw/, scratch/
• Initialize index.md and log.md at root
• Configure the Knowledge Operating System
• Ask if you want to pre-populate with existing knowledge (Memory Integration)
• Become your knowledge maintainer

6. Start using it:

• Drop articles into raw/inbox/ and tell the agent to ingest
• Ask questions and get answers with citations
• Run lint passes to keep the knowledge base healthy

See .kos/PROMPT.md for the full bootstrap prompt and customization options.

Features

Feature	Status
Markdown First	Done
Knowledge Objects (9 types)	Done
YAML Metadata Schema	Done
Agent Workflows (Ingest, Query, Lint, Reflect, Refactor, Promote, Archive)	Done
Memory Integration	Done
Bootstrap from Single Prompt	Done
Obsidian Dataview Queries	Done
Multi-Agent Support	Defined
MCP Integrations	Planned
Knowledge Analytics	Dataview dashboard ready

Repository Structure

llm-knowledge-os/
├── README.md              # This file
├── .kos/                  # Framework (hidden from Obsidian)
│   ├── SCHEMA.md          # The KOS protocol — data format and validation
│   ├── KOS.md             # Instance configuration — domains, access, defaults
│   ├── AGENTS.md          # Agent instructions — workflows, rules, safety
│   ├── PROMPT.md          # Bootstrap prompt — copy, paste, run
│   ├── templates/         # Knowledge Object templates (9 types)
│   ├── obsidian-config/   # Source for Dataview queries and CSS snippets
│   ├── examples/          # Concrete examples of Knowledge Objects
│   └── docs/              # Philosophy, architecture, roadmap
│
│   (After bootstrap, the agent creates at vault root:)
├── .obsidian/             # Obsidian config (copied from .kos/obsidian-config/)
├── knowledge/             # Knowledge Objects (created by agent)
├── raw/                   # Immutable source documents
├── scratch/               # Human-only notepad
├── index.md               # Content catalog
└── log.md                 # Activity log

The .kos/ directory is the framework — it never changes. The agent reads from .kos/ and writes to the vault root. Obsidian sees your knowledge, not the framework.

Objectives

• Human readable
• AI writable
• AI maintainable
• Agent friendly
• Retrieval optimized
• Long-term memory
• Portable
• Markdown only
• Git friendly

Why Another Project?

This project exists thanks to Andrej Karpathy's LLM Wiki.

LLM Wiki introduced the idea of using an LLM as the maintainer of a persistent wiki.

LLM Knowledge OS builds on top of that idea by introducing a lightweight operating layer that makes the wiki easier to understand, maintain and use by autonomous AI agents.

Nothing in LLM Wiki changes. KOS is an optional extension.

The Vision

Imagine opening your knowledge base after six months of AI-assisted maintenance.

You don't find thousands of disconnected notes.

You find an organized knowledge system.

Every concept is connected.

Every project documents its decisions.

Every procedure has been refined over time.

Every meeting has contributed to a growing understanding.

Your knowledge compounds instead of accumulating.

The wiki is no longer a collection of markdown files.

It becomes an operating system for knowledge.

Roadmap

See .kos/docs/roadmap.md for the full roadmap.

Phase	Description	Status
0 — Bootstrap	Single-prompt KOS creation	Done
1 — LLM Wiki Compatibility	Preserve original architecture	Done
2 — Knowledge Objects	Typed, structured knowledge	Done
3 — Metadata Schema	YAML frontmatter with semantic meaning	Done
4 — AI-first Writing Standard	Optimize for LLM consumption	Done
5 — Agent Workflows	Ingest, Query, Lint, Reflect, Refactor, Promote, Archive	Done
6 — Multi-Agent Collaboration	Specialized agents, shared knowledge	Defined
7 — MCP & External Integrations	GitHub, Notion, Gmail, APIs	Planned
8 — Knowledge Analytics	Understand what you know	Partial
9 — Autonomous Knowledge Evolution	Self-improving knowledge system	Planned

License

MIT

Credits

Built on Andrej Karpathy's LLM Wiki.

LLM Wiki organizes knowledge. LLM Knowledge OS organizes how knowledge evolves.