LLM Wiki

Agent-native persistent knowledge management — compile knowledge once, query forever.

Based on Andrej Karpathy’s LLM Wiki pattern.

What is this?

LLM Wiki is a CLI tool + AI Agent skill system that maintains an evolving, interconnected Markdown knowledge base. Instead of traditional RAG (re-deriving answers from raw documents each time), LLM Wiki compiles knowledge into structured wiki pages that AI agents maintain and grow over time.

Key principle: The tool itself doesn’t call LLMs. It provides skill files that let any AI agent (Claude Code, Codex, etc.) operate the wiki. Obsidian is the human interface — no self-built GUI.

Quick Start

# Install globally
npm install -g @jackwener/llm-wiki

# Initialize a new wiki vault
mkdir my-wiki && cd my-wiki
llm-wiki init

# Now use your AI agent:
#   /ingest sources/some-article.md
#   /query "What do we know about X?"
#   /lint
#   /research "deep dive on Y"

llm-wiki init is the only setup command — it creates the vault files, the agent bootstrap files (CLAUDE.md, AGENTS.md), and installs the four bundled operation skills into .claude/skills/ and .agents/skills/ in one step.

After upgrading the package, refresh the installed skill files with:

llm-wiki skill install

Vault Structure

my-wiki/
├── CLAUDE.md              # Agent bootstrap for Claude Code (auto-loaded)
├── AGENTS.md              # Agent bootstrap for Codex (auto-loaded)
├── wiki-purpose.md        # Wiki scope and audience
├── wiki-schema.md         # Page types, naming conventions, frontmatter rules
├── wiki-log.md            # Append-only operation log
├── wiki/                  # AI-maintained wiki pages (Obsidian-compatible)
├── sources/               # Raw, immutable source documents
│   └── YYYY-MM-DD/        # Date-based storage
├── .claude/
│   └── skills/
│       ├── ingest/SKILL.md    # Source → wiki pages
│       ├── query/SKILL.md     # Evidence-grounded wiki answers
│       ├── lint/SKILL.md      # Wiki health checks
│       └── research/SKILL.md  # External research → wiki
├── .agents/
│   └── skills/
│       ├── ingest/SKILL.md
│       ├── query/SKILL.md
│       ├── lint/SKILL.md
│       └── research/SKILL.md
└── .llm-wiki/
    ├── config.toml        # Vault configuration
    └── sync-state.json    # Incremental sync tracking

llm-wiki init generates every file above in one step.

Agent Bootstrap

LLM Wiki uses a two-file pattern so any AI agent can operate the vault out of the box, with no manual setup beyond llm-wiki init:

1. Entry files — CLAUDE.md and AGENTS.md (vault root)

Short bootstrap documents that are auto-loaded on every session start — Claude Code reads CLAUDE.md, Codex reads AGENTS.md. They tell the agent:

• this workspace is an LLM Wiki vault
• where to find wiki-purpose.md and wiki-schema.md
• which /ingest, /query, /lint, /research commands are available
• which of the four operation skills to load (ingest, query, lint, or research)
• a short CLI cheat-sheet and the core operating rules

Because they are auto-loaded, they are intentionally small — a few dozen lines — to keep session-start context cheap.

2. Operation skills — .claude/skills/{ingest,query,lint,research}/SKILL.md and .agents/skills/{ingest,query,lint,research}/SKILL.md

Each operation has a focused playbook, loaded on demand when the matching command is invoked. The skills contain the relevant procedure, page schemas, frontmatter rules, and invariants (for example, source immutability and the wiki-log.md + sync tail rule). Every skill follows the Agent Skills specification (skills//SKILL.md with YAML frontmatter), so spec-compliant agents (Claude Code, Codex, Amp, and others) discover them independently. All four are installed into both platform directories for a shared vault setup.

Upgrading. llm-wiki init is the only setup command — it writes both entry files and installs the four skills. After upgrading the npm package, run llm-wiki skill install to refresh the skill files. Your edits to CLAUDE.md / AGENTS.md are preserved across reinstalls.

Upgrading to the four-skill layout. Run llm-wiki skill install to add the operation skills. If you previously installed the legacy monolithic skill, remove it afterwards so its broad trigger does not overlap the four focused skills:

rm -rf .claude/skills/llm-wiki .agents/skills/llm-wiki

Operations

The skill exposes four operations, each invoked as a slash command:

CLI Commands

Search

BM25 keyword search with CJK bigram tokenization (Chinese/Japanese/Korean support).

When DB9 is configured, search becomes hybrid: BM25 + vector similarity, merged via Reciprocal Rank Fusion (RRF, K=60).

llm-wiki search "distributed consensus"
llm-wiki search "分布式共识" -n 5
llm-wiki search "raft algorithm" --bm25-only

Graph Analysis

Analyzes the [[wikilink]] graph to find structure in your knowledge:

• Communities — Topic clusters detected via label propagation
• Hub pages — Most connected pages (high incoming + outgoing links)
• Orphan pages — Pages with no incoming links
• Wanted pages — Pages linked but not yet created

llm-wiki graph          # Human-readable output
llm-wiki graph --json   # Machine-readable for programmatic use

DB9 Integration (Optional)

DB9 adds vector search and cloud sync:

• Server-side embeddings via embedding(text)::vector(1024) — no local model needed
• HNSW vector index for semantic similarity search
• Reverse source lookup: “which wiki pages reference this source?”

Enable by adding to .llm-wiki/config.toml:

[db9]
url = "your-db9-connection-string"

Then run llm-wiki sync to upload embeddings.

Obsidian Compatibility

The wiki/ directory is a standard Obsidian vault:

• YAML frontmatter
• [[wikilink]] cross-references
• Open directly in Obsidian for browsing, graph view, and editing

Configuration

.llm-wiki/config.toml:

[vault]
name = "My Wiki"
language = "en"

# Optional: DB9 for vector search + cloud sync
# [db9]
# url = "your-db9-connection-string"

Tech Stack

• TypeScript (ESM, Node 20+)
• Commander.js — CLI framework
• gray-matter — Frontmatter parsing
• pg — PostgreSQL client (for DB9)
• tsup — Build
• Vitest — Testing

License

Apache-2.0