LLM Wiki Template

A personal knowledge base that compounds over time — maintained entirely by an LLM agent.

Inspired by Andrej Karpathy's LLM Wiki pattern (April 2026).
Implementation by Bashir Aziz · 2026.

The idea in one paragraph

Most people use LLMs like a search engine — ask a question, get an answer, move on. Nothing accumulates. This template gives your LLM agent a persistent wiki: a directory of markdown files it reads, writes, and cross-references across every session. When you add a new source, the agent doesn't just answer a question — it updates entity pages, revises concept summaries, flags contradictions, and files the knowledge permanently. Every session makes the wiki richer. Every question you ask can become a new wiki page. The knowledge compounds instead of evaporating.

How it works

┌─────────────────────────────────────────────────────────┐
│                        YOU                              │
│   source documents · questions · direction              │
└──────────────────────────┬──────────────────────────────┘
                           │
                           ▼
┌─────────────────────────────────────────────────────────┐
│                    LLM AGENT                            │
│   reads · writes · cross-references · maintains        │
└──────────┬────────────────────────────┬─────────────────┘
           │                            │
           ▼                            ▼
┌──────────────────┐        ┌───────────────────────────┐
│   raw/           │        │   wiki/                   │
│   source docs    │        │   index.md  log.md        │
│   (immutable)    │        │   concepts/ entities/     │
│                  │        │   sources/  threads/      │
└──────────────────┘        └───────────────────────────┘
                                        │
                                        ▼
                            ┌───────────────────────────┐
                            │   sessions/               │
                            │   auto-exported transcripts│
                            │   SQLite FTS5 index       │
                            │   (local, never committed)│
                            └───────────────────────────┘

Three layers:

raw/ — your source documents. Immutable. The LLM reads but never modifies.
wiki/ — LLM-maintained markdown pages. Commits to git. Syncs across devices.
sessions/ — auto-exported session transcripts. Local only. Indexed for search.

Tool support

Tool	Schema file	Auto-export hooks	Quality
Claude Code	`CLAUDE.md`	✅ Full (PreCompact + SessionEnd + SessionStart)	⭐⭐⭐⭐⭐ Best experience
OpenAI Codex CLI	`AGENTS.md`	⚠️ Partial (manual export step)	⭐⭐⭐⭐
Cursor	System prompt file	❌ Manual export	⭐⭐⭐
Aider	`CONVENTIONS.md`	❌ Manual export	⭐⭐⭐
Any LLM via API	Any filename	❌ Manual export	⭐⭐⭐

Claude Code gets the best experience because it's currently the only tool with a native hook system that makes session export fully automatic. With other tools you get ~80% of the value — the wiki compounding, full-text search, Obsidian interface — but session capture requires a manual step.

Windows users: The SessionEnd hook may not fire reliably on exit. See Part 5 of the Setup Guide for the wikiexit PowerShell alias that solves this. Full Windows-specific notes in docs/windows-setup.md.

Quickstart (Claude Code)

# 1. Clone the template
git clone https://github.com/bashiraziz/llm-wiki-template.git my-wiki
cd my-wiki

# 2. Run the setup script
bash scripts/setup.sh

# 3. Create your private wiki repo on GitHub, then:
git remote set-url origin git@github.com:bashiraziz/my-wiki.git
git push -u origin main

# 4. Open a Claude Code session and bootstrap
claude
> customize    # walks you through domain setup interactively
> bootstrap    # seeds all directories and wiki structure

That's it. Your first ingest:

> ingest research raw/research/my-first-article.md

📋 New to this? Start with the Quick Reference Card — a one-page cheat sheet covering every daily workflow command. Print it and keep it handy until the workflow is muscle memory.

Going further

Once your wiki is running, these guides cover the next level:

Guide	What it covers
Quick Reference Card	Daily workflow cheat sheet — print this
Cross-project wiring	Connect all your project repos to the central wiki — one search finds sessions from everywhere
Obsidian setup	Full Obsidian configuration, Web Clipper, graph view tips, mobile setup, Dataview queries
New project checklist	5-minute setup for every new project repo going forward

Repository structure

llm-wiki-template/
│
├── README.md                        ← you are here
├── SETUP-GUIDE.md                   ← complete step-by-step instructions
├── CHANGELOG.md                     ← version history
│
├── scripts/                         ← tool-agnostic shell + python scripts
│   ├── setup.sh                     ← one-time setup script
│   ├── export-session.py            ← session → markdown export
│   ├── index-sessions.sh            ← markdown → SQLite FTS5 indexer
│   ├── recall.sh                    ← search past sessions
│   ├── wire-project.py              ← wire any project to the wiki
│   └── wire-all-projects.py         ← wire all existing projects
│
├── adapters/                        ← tool-specific configuration
│   ├── claude-code/                 ← Claude Code adapter (reference implementation)
│   │   ├── CLAUDE.md                ← schema file (copy to repo root)
│   │   ├── .claude/
│   │   │   └── settings.json        ← hook configuration
│   │   └── README.md                ← Claude Code specific notes
│   ├── codex/
│   │   ├── AGENTS.md                ← schema file for OpenAI Codex
│   │   └── README.md
│   ├── cursor/
│   │   └── README.md
│   └── generic/
│       ├── WIKI-SCHEMA.md           ← tool-agnostic schema template
│       └── README.md
│
├── examples/                        ← filled-in configurations for reference
│   ├── domains/
│   │   ├── govcon-domain.md         ← government contracting domain config
│   │   ├── research-domain.md       ← academic / deep research domain config
│   │   ├── personal-domain.md       ← personal knowledge / journaling domain
│   │   └── business-domain.md       ← business / competitive intelligence domain
│   └── use-cases/
│       ├── solo-researcher.md       ← one person, one domain, simple setup
│       ├── multi-domain.md          ← Bashir's setup: 3 domains, multi-device
│       └── team-wiki.md             ← shared wiki with human review loop
│
├── wiki/                            ← empty wiki structure (ready to populate)
│   ├── index.md
│   └── log.md
│
├── raw/                             ← empty source document directories
│   └── .gitkeep
│
├── sessions/                        ← auto-exported transcripts (gitignored)
│   ├── exports/
│   ├── confidential/
│   └── wiki-digests/
│
├── .gitignore
└── .exportignore

Customization in 3 steps

Step 1 — Choose your adapter

Copy the adapter files for your tool to the repo root:

# For Claude Code:
cp adapters/claude-code/CLAUDE.md .
cp -r adapters/claude-code/.claude .

# For Codex:
cp adapters/codex/AGENTS.md .

# For others: see adapters/generic/README.md

Step 2 — Define your domains

Edit the schema file (CLAUDE.md, AGENTS.md, etc.) and replace the domain placeholders:

YOUR_DOMAIN_1  →  research
YOUR_DOMAIN_2  →  personal
YOUR_DOMAIN_3  →  work

See examples/domains/ for fully worked examples of different domain types.

Step 3 — Bootstrap

claude          # or your tool of choice
> bootstrap     # creates all directories and seed pages for your domains

Multi-device workflow

The wiki pages sync via git. Session transcripts stay local.

# Every session, every device:
git pull && claude

# After every session:
git add wiki/ && git commit -m "session: [topic]" && git push

Full instructions in SETUP-GUIDE.md.

Confidentiality

Three layers of protection for sensitive sessions:

# Layer 1 — Skip export entirely (one command before session):
touch .claude/no-export

# Layer 2 — Export but exclude from search index (.exportignore):
*_client_*.md
*_nda_*.md

# Layer 3 — Encrypt confidential archive (GPG):
# export-session.py --label confidential  →  GPG-encrypted to sessions/confidential/

Known issues and workarounds

Issue	Affects	Workaround
SessionEnd hook doesn't fire on exit	Windows	Use `wikiexit` PowerShell alias — see Setup Guide Part 5
Emoji display as `?` in terminal output	Windows cp1252	Cosmetic only — exports still work correctly
sqlite3 not found in Git Bash	Windows	Copy sqlite3.exe to `C:\Program Files\Git\usr\bin\` as admin

Contributing

PRs welcome, especially:

New tool adapters (Gemini CLI, Aider, Zed, etc.)
New domain examples
Improvements to the export / indexing scripts
Windows compatibility fixes

See CONTRIBUTING.md for guidelines.

License

MIT. Use it, fork it, build on it.

Pattern by Andrej Karpathy · Implementation by Bashir Aziz · Built with Claude (Anthropic) · 2026

Llm Wiki Template