🧠 LLM Wiki Template

An AI-managed personal knowledge base using the Karpathy LLM Wiki Pattern — where LLMs write and maintain a structured Obsidian wiki from your raw research data.

🌐 Language: English | Tiếng Việt

📺 Video Hướng Dẫn

Bộ Nhớ AI Kiểu Karpathy: 3 Bước Xây Wiki Cho Agent [Miễn Phí] — Hướng dẫn chi tiết từ setup đến sử dụng thực tế.

What Is This?

This is a ready-to-use template for building an AI-powered personal knowledge base, inspired by Andrej Karpathy's approach to using LLMs for knowledge management:

"Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest." — Andrej Karpathy

Instead of relying on complex RAG pipelines or vector databases, this system uses a simpler approach:

You dump raw sources (articles, tweets, papers, videos) into raw/
The LLM compiles structured wiki articles in wiki/
You ask questions and get answers grounded in your personal knowledge base
Knowledge compounds — each cycle makes the wiki richer

The result is a 100% inspectable, file-based knowledge system where you can see exactly what your AI "knows."

Key Features

📂 File-based architecture — Markdown files, no databases, no vendor lock-in
🔍 100% inspectable — Every piece of knowledge is a readable .md file
🔄 10 automated workflows — /ingest, /compile, /ask, /cleanup, /breakdown, /autoresearch, /save, /overview, /startup, /wrapup
🔬 Autonomous research — Agent searches the web, evaluates sources, and ingests automatically
⚖️ Contradiction detection — Flags conflicting claims instead of silently overwriting
💾 Chat-to-Wiki pipeline — Save knowledge from conversations directly to wiki
🔌 Integrated MCP Server — Standard MCP API (Model Context Protocol) with FTS5 search for AI agents (Secure 127.0.0.1 bind)
🕸️ Knowledge Graph — Automated analysis and visualization of knowledge links (God nodes, Orphans)
📊 Self-maintaining indexes — Master index, glossary, backlinks, executive overview, operations log
🛡️ Quality gates — Article size guardrails, anti-cramming/thinning rules, re-read checks
🧹 Wiki health checks — Automated tone, structure, link, and contradiction auditing
📈 Compound knowledge loop — Each cycle produces better knowledge, which produces better outputs

Quick Start

1. Use This Template

Click "Use this template" → "Create a new repository" on GitHub.

Or clone manually:

git clone https://github.com/YOUR_USERNAME/llm-wiki-template.git my-second-brain

2. Open in Obsidian

Download Obsidian (free)
Open as vault: File → Open vault → Open folder as vault
Select the cloned directory
Install recommended plugins when prompted: Dataview, Marp Slides

3. Connect Your AI Agent

This template works with any LLM-powered coding agent that can read files. Tested with:

Gemini CLI (recommended)
Claude Code / Claude Desktop with filesystem access
Cursor / Windsurf with workspace access
Any agent that can read/write Markdown files

The agent reads AGENTS.md as its operating manual — no additional configuration needed.

4. Start Building Your Knowledge Base

# Step 1: Ingest a source
/ingest https://example.com/interesting-article

# Step 2: Compile into wiki
/compile

# Step 3: Ask questions
/ask What are the key concepts from my sources?

# Step 4: Audit wiki quality
/cleanup

# Step 5: Find knowledge gaps
/breakdown

# Step 6: Auto-research a topic
/autoresearch Large Language Models

# Step 7: Save chat insights to wiki
/save

Architecture

┌─────────────────────────────────────────────────┐
│                 YOUR RESEARCH                    │
│  Articles, Tweets, Papers, Videos, Repos, etc.   │
└──────────────────────┬──────────────────────────┘
                       │ /ingest
                       ▼
┌─────────────────────────────────────────────────┐
│                  raw/                            │
│  Source documents — NEVER modified, only added    │
│  articles/ papers/ repos/ tweets/ videos/ misc/  │
└──────────┬───────────────────────┬──────────────┘
           │ /compile              │ /autoresearch
           ▼                      ▼
┌─────────────────────────────────────────────────┐
│                  wiki/                           │
│  Compiled knowledge — AI-maintained wiki          │
│  concepts/ tools/ people/ comparisons/           │
│  + _index.md, _glossary.md, overview.md          │
│  ⚖️ Contradiction Check before every update      │
└──────┬─────────┬─────────┬──────────────────────┘
       │ /ask    │ /cleanup │ /save
       ▼        ▼          ▼
┌──────────┐ ┌──────────┐ ┌──────────────────────┐
│ answers  │ │ quality  │ │ chat → raw → wiki     │
│ + refs   │ │ fixes    │ │ knowledge extraction  │
└──────────┘ └──────────┘ └──────────────────────┘

Workflows

Command	What It Does
`/ingest`	Imports raw sources (URLs, files, PDFs) into `raw/` with proper frontmatter
`/compile`	Reads raw sources and creates/updates structured wiki articles (with contradiction detection)
`/ask`	Answers questions using wiki knowledge, with optional file-back to wiki
`/cleanup`	Audits wiki quality — tone, structure, links, size, contradiction backlog — and auto-fixes
`/breakdown`	Scans wiki for missing entities and proposes new articles
`/autoresearch`	🆕 Autonomous research — searches the web, evaluates sources, ingests, and synthesizes reports
`/save`	🆕 Chat-to-Wiki — extracts knowledge from conversations and saves directly to wiki
`/startup`	🆕 Project Brain Startup — AI recalls context from previous sessions
`/wrapup`	🆕 Project Wrapup — AI saves session archive and updates rolling context

Each workflow is defined in .agents/workflows/ and can be customized.

AutoResearch — Autonomous Knowledge Discovery

The /autoresearch workflow turns your wiki into an active researcher:

/autoresearch [topic]

How it works:

Gap Analysis — Scans existing wiki to identify what's missing
3-Round Research Loop — Broad search → Gap fill → Verify
Auto-Ingest — Downloads and processes sources automatically
Synthesis Report — Generates an executive summary at outputs/reports/
Human Review — You approve before anything enters the wiki

Configure search constraints in raw/_research_program.md.

Contradiction Detection

When compiling new sources, the system automatically checks for conflicting claims:

✅ Temporal updates (v1.0 → v2.0) — Updated normally
✅ New information — Integrated normally
⚠️ Actual contradictions — Preserved with [!warning] callout, tagged needs-review

The wiki never silently overwrites conflicting information. Human review is always required.

Quality System

The template enforces several quality mechanisms:

Re-read before update — The AI must read the full article before editing (non-negotiable)
Contradiction check — Compare new claims against existing wiki before writing
Article size guardrails — 15–120 lines; too short = stub, too long = split
Anti-cramming — Sub-topics with ≥3 paragraphs get their own article
Anti-thinning — No article creation unless ≥3 meaningful sentences can be written
Encyclopedia tone — Neutral, attribution-based writing, no editorial voice
Absorption log — Tracks which raw sources have been compiled (no duplicates)
Operations log — Chronological record of every action taken on the vault

File Structure

llm-wiki-template/
├── AGENTS.md                ← Agent operating manual (the brain)
├── README.md                ← This file
├── update.py                ← 🆕 One-command updater script
├── sync-brain.ps1           ← 🆕 Auto-sync to GitHub script
├── .gitignore
│
├── .agents/workflows/       ← 10 automated workflows
│   ├── ask.md
│   ├── autoresearch.md      ← 🆕 Autonomous research
│   ├── breakdown.md
│   ├── cleanup.md           ← Updated: contradiction backlog scanning
│   ├── compile.md           ← Updated: contradiction detection (Step 4.5)
│   ├── ingest.md
│   ├── save.md              ← 🆕 Chat-to-Wiki pipeline
│   ├── startup.md           ← 🆕 Session startup
│   └── wrapup.md            ← 🆕 Session wrapup
│
├── integrations/mcp/        ← 🆕 MCP Server Integration
│   ├── README.md            ← MCP Setup Guide
│   └── config-sample.json   ← Sample config for Claude/Cursor
│
├── scripts/                 ← 🆕 Agent-Native Tooling
│   ├── brain.py             ← Central CLI Router (Search, Index, Health, MCP)
│   ├── brain_mcp.py         ← FastMCP Server (Secure 127.0.0.1 bind)
│   ├── brain_db.py          ← SQLite Database abstraction
│   ├── build_search_index.py← FTS5 Indexing
│   └── ...                  ← Other tools (audit, resolve_orphans, etc.)
│
├── .obsidian/               ← Obsidian config (pre-configured)
│
├── sessions/                ← Session logs (AI Memory)
│   ├── current-context.md   ← Rolling context (auto-updated)
│   ├── .hot-buffer.md       ← Mid-session decisions buffer
│   └── session-summary-*.md ← Archive of each session
│
├── raw/                     ← Your source documents
│   ├── _ingest.py           ← Batch ingest script (Python)
│   ├── _research_program.md ← 🆕 AutoResearch configuration
│   ├── articles/
│   ├── papers/
│   ├── repos/
│   ├── tweets/
│   ├── videos/
│   └── misc/
│
├── wiki/                    ← AI-maintained wiki
│   ├── overview.md          ← 🆕 Executive summary for cross-project access
│   ├── _index.md            ← Master catalog
│   ├── _glossary.md         ← Term definitions
│   ├── _absorb_log.json     ← Compilation tracker
│   ├── _backlinks.json      ← Reverse link index
│   ├── _build_backlinks.py  ← Backlinks builder script
│   ├── _build_graph.py      ← 🆕 Knowledge Graph analysis script
│   ├── _dashboard.md        ← Dataview dashboard
│   ├── _ops_log.md          ← Operations log
│   ├── concepts/
│   ├── tools/
│   ├── people/
│   └── comparisons/
│
└── outputs/                 ← Generated content
    ├── reports/             ← AutoResearch synthesis reports
    ├── slides/
    ├── charts/
    └── summaries/

Customization

Change the Language

The template uses English by default. To switch:

Edit AGENTS.md → update the Writing Tone section
Update wiki meta files (_index.md, _glossary.md) headers
The AI will follow your language preference from AGENTS.md

Add Entity Types

Edit AGENTS.md → Entity-Type Templates section to add new categories beyond concepts/tools/people/comparisons.

Modify Quality Rules

All quality rules are in AGENTS.md. Adjust thresholds (article size, quote density, etc.) to match your preferences.

Configure AutoResearch

Edit raw/_research_program.md to customize:

Search scope and constraints
Confidence scoring thresholds
Source exclusion lists
Domain-specific notes and priorities

Add Obsidian Plugins

The template includes configs for Dataview (tables/queries) and Marp Slides (presentations). Add more plugins through Obsidian's community plugin browser.

Batch Ingest Script

For bulk importing, use the included Python script:

# Single file
python raw/_ingest.py path/to/article.md

# PDF (requires PyMuPDF: pip install PyMuPDF)
python raw/_ingest.py paper.pdf

# Entire folder
python raw/_ingest.py ~/Downloads/research-notes/

# Preview without creating files
python raw/_ingest.py big-folder/ --dry-run

Updating Your Vault

Already using the template? Update to the latest workflows with a single command:

Option A: `brain` CLI (Recommended)

Install once:

pip install git+https://github.com/KHOAAI-HILL/llm-wiki-template.git

Then use anywhere inside your vault:

brain update              # Update (asks for confirmation)
brain update --dry-run    # Preview changes without writing
brain update --force      # Update without asking
brain status              # Check vault health
brain version             # Show version

Option B: Standalone script (No install)

# Download the script
curl -o update.py https://raw.githubusercontent.com/KHOAAI-HILL/llm-wiki-template/master/update.py

# Run it
python update.py

Both methods only touch system files (workflows, AGENTS.md). Your personal data (raw/, wiki/, sessions/, outputs/) is never modified.

Philosophy

This template is built on three principles:

Files over databases — Markdown files are portable, inspectable, and version-controllable. No vector DB, no cloud dependencies.
Compile once, query forever — Instead of retrieving raw chunks on every query (RAG), the AI pre-compiles clean wiki articles. Queries read refined knowledge, not raw data.
Knowledge compounds — Each ingest-compile-ask cycle makes the wiki richer. Better wiki → better answers → better questions → richer wiki.

Credits

Andrej Karpathy — Originated the LLM Knowledge Base concept
Farzaa — wiki-gen-skill implementation that heavily influenced quality gates
DataChaz — Community breakdown and analysis

License

MIT — Use freely, modify as you wish, share with others.

Llm Wiki Template

About Llm Wiki Template

Platforms

Languages

Links

Need Help Installing Llm Wiki Template?