The Problem
LLMs lose all context between sessions. RAG systems re-derive knowledge from raw documents on every query, spending tokens to rediscover what was already known. In 2025, Andrej Karpathy proposed a different approach: have the LLM incrementally build and maintain a persistent wiki -- structured, interlinked markdown that compounds over time. The idea sparked a 400+ comment discussion drawing on the Zettelkasten tradition and digital garden movement.
This plugin turns that idea into a working system. It is a Claude Code plugin with 4 task-oriented agents, 7 internal workflow skills, 2 slash commands, and 27 utility scripts that automate the full capture / research / answer / audit / maintain loop inside an Obsidian vault. Knowledge is compiled once and kept current -- not re-derived on every query.
What's new in v0.4.0
- Agent-first UX — talk to the system in natural language; Claude triggers the right agent
- Extended personas with voice / vocabulary / anti-patterns / litmus tests
- Sonnet-only for token economy
- defuddle → trafilatura → r.jina.ai privacy-first web→md stack (cloud opt-in)
- hot.md rolling 500-word session cache
- Tree topology lints — index ≤100 links, hubs ≤15 members
- Confidence + supersession in frontmatter
- Claim-level citations with raw quotes and anchors
- Forgetting curve for stale-page detection
[!IMPORTANT]
Trust and transparency
- Local-only: All data stays in your Obsidian vault on your machine. No cloud sync, no external storage.
- No telemetry: No analytics, no usage tracking, no phone-home behavior of any kind.
- Network access: Only capture scripts (
capture-url.sh, capture-youtube.sh, capture-github.sh) make network requests, and only to URLs you explicitly provide.
- What it creates: Files in
raw/, wiki/, index.md, log.md, wiki.config.md, and .obsidian/ config -- all inside your vault.
- Uninstall:
/plugin remove llm-obsidian-wiki@ignromanov -- your vault data remains untouched.
Install
Prerequisites: Claude Code, Obsidian, Obsidian CLI
From the Claude Code plugin marketplace:
# Add the marketplace source
/plugin marketplace add ignromanov/llm-obsidian-wiki
# Install the plugin
/plugin install llm-obsidian-wiki@ignromanov
Or load directly from a local clone:
git clone https://github.com/ignromanov/llm-obsidian-wiki.git
claude --plugin-dir /path/to/llm-obsidian-wiki
See It Work
$ claude
> /wiki:init
What is the project/wiki name? ml-research
Which page types? (all 10 selected)
Which raw source directories? (defaults: knowledge, specs, meetings, docs, external, inbox)
Which capture tools are installed? defuddle, trafilatura
Wiki vault "ml-research" created successfully.
Files: CLAUDE.md, wiki.config.md, index.md, hot.md, log.md
Next: open the vault in Obsidian, add your first source.
> "research karpathy llm wiki pattern"
[wiki-researcher]
Capturing: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f
Saved: raw/external/2026-04-11-karpathy-llm-wiki-pattern.md
Compiling wiki pages...
Created: wiki/sources/src-karpathy-llm-wiki-pattern.md
Created: wiki/concepts/llm-wiki-pattern.md (confidence: high, 3 sources)
Created: wiki/concepts/compounding-knowledge.md
Updated: index.md, hot.md, log.md
> "how does compounding knowledge differ from RAG?"
[wiki-advisor]
Searching wiki... 3 pages found, reading 2...
RAG re-derives answers from raw documents on each query, paying full token
cost every time. The LLM Wiki pattern ([[llm-wiki-pattern]]) compiles
knowledge once into structured [[compounding-knowledge]] pages maintained
incrementally. Subsequent queries read the compiled wiki instead of
re-processing raw sources.
Sources: [[src-karpathy-llm-wiki-pattern]], [[compounding-knowledge]]
> "clean up the wiki"
[wiki-curator]
Running health checks...
✓ No orphaned pages
✓ No broken wikilinks
⚠ 1 stale page detected: wiki/concepts/compounding-knowledge.md (90 days)
Auto-fixed: index drift resolved
Report saved: log.md
Quick Start
kfchou
