Hermes LLM Wiki: What Karpathy's Wiki Skill Does, How To Set It Up, And The Models To Run It On

The Hermes LLM wiki is the bundled llm-wiki skill in Hermes Agent, which makes the agent build and maintain a persistent knowledge base as a folder of interlinked markdown pages.

It is based on the LLM Wiki pattern that Andrej Karpathy published as a GitHub gist in April 2026, and Nous Research ships it with Hermes, installed by default.

You set it up by pointing the WIKI_PATH variable at a folder and typing /llm-wiki, and Hermes writes the schema, index and log for you.

I'm Julian Goldie, and I'll explain what the skill does, how to set it up and how each operation works, and then I'll show you which models to run it on.

Hermes LLM Wiki at a glance

This is the whole thing on one screen, taken from the official docs.

What is the Hermes LLM Wiki?

Hermes Agent is the free, open-source AI agent from Nous Research, released under the MIT licence.

The Hermes LLM wiki is one of its bundled skills, and the official Llm Wiki docs page describes it as "Karpathy's LLM Wiki: build/query interlinked markdown KB."

A skill is a set of instructions Hermes loads when a task matches, so the wiki is a disciplined workflow rather than a separate app.

When I checked my own Hermes install, which runs v0.22.1, the skill was already in my skills folder at version 2.1.0.

The skill is free, so the only running cost is the model you connect to Hermes.

Karpathy's LLM Wiki idea in plain English

Karpathy's gist starts from a simple complaint about how most people use AI with documents.

Tools built on RAG pull relevant chunks at question time, so the model rediscovers the same knowledge every time you ask.

His alternative is an agent that reads each source once and integrates it into a wiki that sits between you and your raw files.

In his words, "The knowledge is compiled once and then kept current, not re-derived on every query."

He wrote the gist as an idea file for agents such as Codex, Claude Code or OpenCode, and he deliberately left the details abstract.

The Hermes skill fills in those details with a fixed folder layout, a schema template and step-by-step operations.

The git history of the hermes-agent repo shows the skill landed in April 2026, two days after the gist.

How the Hermes skill differs from Karpathy's original

The repo history shows the provenance markers, source hashing and quality signals were ported into the skill from llm-wiki-compiler a couple of weeks after launch.

That's why my launch-time video covers the core idea and the three operations but not those newer details.

The problem the Hermes LLM Wiki solves

Every research chat with an agent disappears into history once the session ends.

The next week you ask a related question, and the agent starts from zero again.

The Hermes LLM wiki turns each session's work into pages that stay, so the knowledge compounds instead of evaporating.

The docs describe the split of work clearly, because the human curates sources and directs analysis while the agent summarises, cross-references, files and maintains consistency.

How to set up the Hermes LLM Wiki

First, update Hermes, because the docs say each hermes update syncs the bundled skills into ~/.hermes/skills/.

Second, set WIKI_PATH in the .env file in your Hermes home, or leave it unset and the wiki goes in ./wiki.

Third, type /llm-wiki, because the docs say every installed skill is automatically available as a slash command.

You can also preload it for a session, because hermes chat takes -s or --skills to preload skills.

On a new wiki, the agent creates the folders, asks what domain the wiki covers, writes SCHEMA.md, index.md and log.md, and suggests your first sources.

This is the folder structure it creates, copied from the docs.

wiki/
├── SCHEMA.md           # Conventions, structure rules, domain config
├── index.md            # Sectioned content catalog with one-line summaries
├── log.md              # Chronological action log (append-only, rotated yearly)
├── raw/                # Layer 1: Immutable source material
│   ├── articles/       # Web articles, clippings
│   ├── papers/         # PDFs, arxiv papers
│   ├── transcripts/    # Meeting notes, interviews
│   └── assets/         # Images, diagrams referenced by sources
├── entities/           # Layer 2: Entity pages (people, orgs, products, models)
├── concepts/           # Layer 2: Concept/topic pages
├── comparisons/        # Layer 2: Side-by-side analyses
└── queries/            # Layer 2: Filed query results worth keeping

The raw/ layer holds your sources, which the agent reads but never edits.

The agent owns entities/, concepts/, comparisons/ and queries/.

SCHEMA.md is the third layer, and it sets filenames, frontmatter, page thresholds and a taxonomy of 10 to 20 tags.

At the start of every session the agent reads SCHEMA.md, index.md and the last 20 to 30 log entries before doing anything else.

How to use the Hermes LLM Wiki: ingest, query and lint

Ingest asks for your approval first if it would touch 10 or more existing pages.

The hash on each raw file means a re-ingest skips an unchanged source and flags drift when the source has changed.

Pages that synthesise three or more sources carry provenance markers such as ^[raw/articles/source-file.md], so every claim can be traced.

Lint looks for orphan pages, broken wikilinks, pages missing from the index, missing frontmatter, stale pages, contradictions, low-confidence claims and source drift.

It also flags pages over 200 lines for splitting and rotates log.md once it passes 500 entries.

The optional confidence, contested and contradictions frontmatter fields are what lint uses to surface weak claims.

Every other Hermes LLM Wiki feature

Bulk ingest reads all the sources first, checks existing pages in one pass and writes a single log entry for the batch.

Archiving moves a superseded page to _archive/, removes it from the index and replaces links to it with plain text marked as archived.

The index splits a section once it passes 50 entries, and a _meta/topic-map.md appears once the index passes 200 entries.

The docs say the folder works as an Obsidian vault out of the box, with clickable wikilinks, Graph View and Dataview queries.

They recommend setting Obsidian's attachment folder to raw/assets/, and pointing OBSIDIAN_VAULT_PATH at the wiki if you also use the obsidian skill.

On a server, the skill documents npm install -g obsidian-headless, which needs Node.js 22 or newer and an Obsidian Sync subscription, with ob sync --continuous keeping the vault in sync.

Hermes cron jobs can attach skills with the --skill flag, and the skill skips its discuss-takeaways step in cron contexts, so scheduled ingests fit naturally.

For a wider view of how Hermes remembers things, my Hermes Agent memory roundup compares the options.

Hermes LLM Wiki troubleshooting

The docs also mention llm-wiki-compiler, a Node.js CLI for batch-compiling a folder into a wiki, which they say is tuned for small corpora.

The limits of the Hermes LLM Wiki

Karpathy says an index file works well at around 100 sources and hundreds of pages, and the skill adds file search once a wiki passes 100 pages.

Big ingests pause for approval, which is safe but slows fully unattended runs.

Sources sit on disk as plain files, so you choose where the folder lives and who can reach it.

The biggest limit is the one the rest of this page is about, because the skill is only as careful as the model behind it.

The Hermes LLM Wiki doesn't ship its own model

The skill is a set of instructions, so every summary, cross-reference and contradiction flag is written by whichever model you connect to Hermes.

GoldieBench doesn't test wiki maintenance, because it scores one-shot builds rather than long research sessions.

What it does show is how well a model follows a long, detailed brief without dropping steps, which is exactly what a 13-step lint or a 15-page ingest demands.

Treat these numbers as a guide to instruction-following under pressure, not as a wiki benchmark.

Every score above matches the live model pages on goldiebench.com as of 27 October 2026.

If privacy matters more than polish, my local models leaderboard shows what runs on your own machine.

For the full picture of which brain to give Hermes, my best Hermes Agent model breakdown goes deeper.

Who the Hermes LLM Wiki is for

It suits anyone who researches the same subject over weeks or months, including content creators, agencies tracking competitors and teams that want an internal wiki.

In my video I walk through Karpathy's idea, the three layers and the three operations, and I cover researchers, creators, developers and business teams as the people who benefit most.

It is less useful for one-off questions, where a normal chat is faster.

Also On Our Network

• Agent Operatorsthe seven-step Hermes LLM wiki setup
• AI Income Desksix ways a business can use the Hermes LLM wiki
• AI Tool Verdictthe Hermes LLM wiki reviewed and scored
• agentos.guidethe Hermes LLM wiki quick start for an Agent OS
• agentos.guidemy Hermes second brain guide on agentos.guide

If you run an SEO business and want to talk through putting an agent knowledge base to work, you can book a free SEO strategy session with my team.

If you'd like help choosing which model should power your agency's research, my team can cover that in a free SEO strategy session too.

Set the folder, pick a strong model from the live board and lint it every week, and you'll get the most out of the Hermes LLM wiki.