The Hermes LLM wiki is llm-wiki, a skill that ships inside Hermes Agent and turns the agent into the full-time editor of a markdown knowledge base. You feed it sources, it writes and links the pages, and every new article or question makes the wiki smarter. It's Andrej Karpathy's LLM Wiki idea, built straight into Nous Research's agent.
📺 Watch: Hermes AI + Kaparthy LLM Wiki is INSANE!
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside
No plugin to hunt down and no repo to clone. It's bundled and installed by default. Here's what it does, the three jobs it runs, how to set it up, and how to point it at Obsidian.
What is the Hermes LLM wiki?
Start with Karpathy. His LLM Wiki gist calls itself an "idea file", a page you paste into your own agent so it can build the details with you. The core point is simple. Most tools treat your documents like RAG: upload files, pull chunks at question time, answer, forget. Karpathy lists NotebookLM and ChatGPT file uploads as examples. Nothing builds up.
His alternative is a wiki the LLM keeps growing. When you add a source, the agent reads it, pulls out what matters and folds it into pages that already exist. It updates entity pages, rewrites summaries and flags where new data clashes with old claims. He calls the result a persistent, compounding artifact.
Hermes turned that idea into a working skill. Per the Hermes docs, llm-wiki sits in the research category, version 2.1.0, MIT licensed, and runs on Linux, macOS and Windows. Its related skills are obsidian and arxiv. The split of work is clear in the skill file: you curate sources and steer the analysis, the agent summarises, cross-references, files and keeps everything consistent.
I run Hermes daily inside my Agent OS, with a profile per model, and a self-maintaining knowledge base is one of the most useful things you can bolt onto it. If you want the full setup, prompts and coaching, the AI Profit Boardroom ($69/mo) gives you my Agent OS, a 30-day AI agent roadmap, daily tutorials as new tools drop, 4 coaching calls a week and the prompt library. 3,400+ members are already inside.
How the wiki is built: three layers
The skill uses the same three layers Karpathy describes.
- Raw sources. The raw folder holds articles, papers, transcripts and assets. The agent reads these and never edits them. They're your source of truth.
- The wiki. Agent-owned pages in entities, concepts, comparisons and queries folders. The agent creates them, updates them and links them.
- The schema. SCHEMA.md sets the domain, file naming, frontmatter fields, the tag list and the rules for when a page gets created, split or archived.
Two more files keep it navigable. index.md is the catalog, every page with a one-line summary. log.md is an append-only diary of every ingest, query and lint. Karpathy notes that an index like this works well up to roughly 100 sources and hundreds of pages, without any embedding setup.
There's one rule the skill marks as critical. At the start of every session, Hermes reads SCHEMA.md, then index.md, then the last 20 to 30 log entries before it touches anything.
Ingest, query, lint: the three jobs
Ingest
Hand Hermes a URL, a PDF or pasted text. It saves the raw copy with a small header: the source URL, the ingest date and a sha256 hash of the content. Re-ingest the same URL later and it skips the work if nothing changed, or flags the drift if it did. Then it talks through the takeaways with you (it skips that step in cron runs), checks which pages already exist, and writes or updates them. The skill says one source can touch 5 to 15 pages. If an ingest would edit 10 or more existing pages, it asks you first.
Query
Ask a question and Hermes reads the index, opens the relevant pages and answers with citations to the pages it used. The clever part: if the answer is a real piece of work, like a comparison or a deep dive, it files it back as a new page in queries or comparisons. Your questions grow the wiki too.
Lint
Ask for a health check and the skill runs a long list: orphan pages with no inbound links, broken wikilinks, pages missing from the index, bad frontmatter, stale pages, contradictions, low-confidence claims, raw files whose hash changed, pages over about 200 lines, and tags outside the schema. It reports by severity, broken links first.
Want a second pair of eyes on how a knowledge base like this fits your business? Grab a free AI strategy session.
📺 Watch: Hermes3D Is INSANE: AI Agents Now Have a 3D Office
How to set up the Hermes LLM wiki skill
Because it's bundled, setup is mostly telling Hermes where the wiki lives.
- Check the skill is there. Bundled skills are copied into ~/.hermes/skills/ when you install Hermes and synced again on hermes update. If you deleted it at some point, the docs say to restore it with hermes skills reset llm-wiki --restore.
- Pick a folder. Set the WIKI_PATH environment variable, for example in ~/.hermes/.env. Leave it unset and the skill defaults to ~/wiki.
- Start the wiki. Every installed skill becomes a slash command, so type /llm-wiki followed by your request, or just ask Hermes to start a wiki. It also picks the skill on its own when you mention building a knowledge base.
- Name your domain. Hermes asks what the wiki covers. Be specific. It then writes SCHEMA.md, index.md and log.md, builds the folders and suggests your first sources.
- Feed it. Drop in a link or a file and say ingest. Read what it changed, then keep going.
For more on giving Hermes a long-term brain, my Hermes second brain guide covers the wider setup.
📺 Watch: Hermes Bot Mode DESTROYS Grok Bot? (FREE!)
Using the LLM wiki with an Obsidian vault
This is where it gets good. Karpathy describes running the agent on one side of the screen and Obsidian on the other, so you watch pages appear as you talk. The skill says the wiki folder works as an Obsidian vault out of the box.
- Wikilinks. Pages link with double-bracket wikilinks, so they're clickable in Obsidian.
- Graph View. You see the whole network, which pages are hubs and which are orphans.
- Dataview. The YAML frontmatter on every page means the Dataview plugin can build tables, say every entity tagged company.
- Images. Set Obsidian's attachment folder to raw/assets so clipped images land where the agent can find them.
If you also use the bundled obsidian skill, set OBSIDIAN_VAULT_PATH to the same folder as WIKI_PATH so both skills write to one vault. That skill falls back to ~/Documents/Obsidian Vault when the variable isn't set. My Hermes Obsidian walkthrough covers that skill on its own, and Hermes Agent Obsidian memory shows how to use a vault as the agent's memory.
Running Hermes on a server with no screen? The skill documents obsidian-headless, which syncs the vault through Obsidian Sync without the desktop app. It needs Node.js 22 or later and an Obsidian account with a Sync subscription. The steps it lists:
- npm install -g obsidian-headless
- ob login with your Obsidian email and password
- ob sync-create-remote --name "LLM Wiki"
- From inside the wiki folder, ob sync-setup --vault followed by your vault ID
- ob sync for the first pass, then ob sync --continuous to keep it live
The skill also gives a systemd user service so the sync survives logout. Hermes writes on the server, you read the same vault on your laptop or phone.
Hermes LLM wiki vs Karpathy's original idea
| Part | Karpathy's LLM Wiki gist | Hermes llm-wiki skill |
|---|---|---|
| What it is | An abstract idea file you paste into any agent | A bundled skill, installed by default |
| Schema file | Something like CLAUDE.md or AGENTS.md | SCHEMA.md with a ready-made template |
| Folder layout | Left to you and your agent | raw, entities, concepts, comparisons, queries |
| Location | Anywhere | WIKI_PATH, default ~/wiki |
| Operations | Ingest, query, lint | Ingest, query, lint, plus bulk ingest and archiving |
| Change tracking | Not specified | sha256 hash on every raw source |
| Quality flags | Not specified | confidence, contested and contradictions fields |
| Obsidian | Recommended as the viewer | Works as a vault, headless sync documented |
The gist says itself that it's intentionally abstract. The Hermes skill is the opinionated version, with the rules already written.
Tips to keep your wiki healthy
- Ingest one source at a time at first. Karpathy prefers this too.
- Lint on a schedule. A weekly lint catches orphans and contradictions before they pile up.
- Keep the tag list tight. The skill only allows tags defined in SCHEMA.md. Add a tag there first, then use it.
Which model should run your wiki? Check Goldie Bench for how models do on real agent work, and read the Agent OS guide to see how a wiki fits next to your other agents. If you'd like help planning it, book a free AI strategy session.
FAQ
Is there a Hermes LLM wiki plugin I need to install?
No. llm-wiki is a bundled skill, not a plugin. It's installed by default with Hermes Agent. If it's missing, hermes skills reset llm-wiki --restore brings it back.
Where does the wiki get saved?
Wherever WIKI_PATH points. If you never set it, the skill uses ~/wiki. It's a plain folder of markdown files with no database.
Can I use my existing Obsidian vault?
Yes. Point WIKI_PATH at the vault, or a folder inside it, and set OBSIDIAN_VAULT_PATH to match if you use the obsidian skill too. Wikilinks, Graph View and Dataview all work.
How is this different from RAG?
RAG digs through raw files every time you ask. The wiki compiles knowledge once, keeps the cross-references and flagged contradictions, and builds on them with every source.
Ready to set up your Hermes LLM wiki properly? Join the AI Profit Boardroom for the Agent OS and step-by-step tutorials, or book a free AI strategy session and we'll map out your knowledge base together.
