The Hermes LLM wiki is the llm-wiki skill that ships with Hermes Agent, and it makes the agent build and maintain a persistent knowledge base of interlinked markdown pages for you.
It is Nous Research's take on the LLM Wiki pattern that Andrej Karpathy published as a GitHub gist in April 2026, and the official docs say it is installed by default.
You set it up by choosing a folder with the WIKI_PATH variable and typing /llm-wiki in Hermes, and the agent writes the schema, index and log for you.
My verdict is 4.5 out of 5, because it turns an abstract idea into a disciplined, well-documented workflow, and its only real weakness is that it depends on the model you connect.
Below I score every feature against what the official docs say it does, then compare it with RAG and with Karpathy's original.
In the video below I walk through Karpathy's idea, the three layers and the three core operations, which makes the feature breakdown easier to follow.
Hermes LLM Wiki Scores At A Glance
These are my scores for each feature, judged against what the official docs say it does.
| Feature | What it does | My score |
|---|---|---|
| Setup | You set WIKI_PATH and type /llm-wiki, and the agent builds the scaffolding. |
9/10 |
| Plain markdown storage | It keeps everything in a folder of markdown files with no database. | 10/10 |
| Three-layer structure | It separates raw sources, agent-written pages and the schema. | 9/10 |
| Ingest | It captures a source, updates 5 to 15 pages and logs every change. | 9/10 |
| Query | It answers from the wiki with [[wikilink]] citations and files good answers back. |
9/10 |
| Lint | It runs a 13-step health check grouped by severity. | 10/10 |
| Quality signals | It tracks confidence, contested claims and contradictions in frontmatter. | 9/10 |
| Source hashing | It stores a sha256 per raw file to skip unchanged sources and flag drift. |
9/10 |
| Obsidian support | It works as a vault out of the box, with Headless sync for servers. | 9/10 |
| Model dependence | It is only as careful as the model you connect. | 6/10 |
| Overall | It is the most complete free version of Karpathy's pattern I've seen documented. | 4.5/5 |
The full skill definition is on the official Llm Wiki docs page, which reproduces the complete SKILL.md.
🔥 Want my exact Hermes LLM Wiki Review setup, step by step? Inside the AI Profit Boardroom you get step-by-step video trainings, my installable Agent OS and four coaching calls a week with 3,400+ members building real automations. → Join the AI Profit Boardroom
What The Hermes LLM Wiki Is
Hermes Agent is the free, open-source AI agent from Nous Research.
The Hermes LLM wiki is one of its bundled skills, and the docs describe 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 this is a workflow the agent follows rather than a separate app.
| Detail | What the docs list |
|---|---|
| Source | It is bundled and installed by default. |
| Path | It lives at skills/research/llm-wiki. |
| Version | The current version is 2.1.0. |
| Licence | It is released under the MIT licence. |
| Platforms | It runs on Linux, macOS and Windows. |
| Related skills | The related skills are obsidian and arxiv. |
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.
Where The Idea Comes From
Karpathy's LLM Wiki gist argues that tools like NotebookLM and most RAG systems rediscover knowledge from scratch on every question.
His alternative is an agent that incrementally builds and maintains a wiki between you and your sources, and his key line is that "the wiki is a persistent, compounding artifact."
He designed the gist as an idea file to paste into an agent such as Codex, Claude Code or OpenCode, and he says it is intentionally abstract.
The Hermes skill links straight to that gist and fills in the specifics.
How To Set It Up
- You update Hermes with
hermes update, because the docs say every update syncs bundled skills into~/.hermes/skills/. - You set
WIKI_PATHin the.envfile in your Hermes home, or you leave it unset and the wiki goes in~/wiki. - You type
/llm-wiki, because every installed skill is available as a slash command. - You tell the agent what domain the wiki covers, and it writes
SCHEMA.md,index.mdandlog.md. - You give it a first source and watch the first pages appear.
You don't strictly need the slash command, because the skill activates when you ask to create a wiki, ingest a source, lint a wiki or ask a question while a wiki exists.
Every Feature, Reviewed
Storage: just a folder of markdown
The docs say the wiki is "just a directory of markdown files," with no database and no special tooling.
That is the right call, because you can read, edit, back up or version-control the whole thing with tools you already have.
Karpathy makes the same point in his tips, noting that the wiki is just a git repo of markdown files.
The three-layer structure
The folder has a raw/ layer with articles/, papers/, transcripts/ and assets/, which the agent reads but never modifies.
It has an agent-owned wiki layer made of entities/, concepts/, comparisons/ and queries/.
It has a schema layer, SCHEMA.md, which defines the structure, conventions and tag taxonomy.
Karpathy suggests a file like CLAUDE.md or AGENTS.md for the schema, and Hermes uses a dedicated SCHEMA.md instead, which keeps the wiki rules separate from your agent's general instructions.
Session orientation
The skill marks one rule as critical, which is that the agent reads SCHEMA.md, index.md and the last 20 to 30 entries of log.md before doing anything.
The docs say this prevents duplicate pages, missed cross-references and repeated work.
On wikis with 100 or more pages, the agent also searches for the topic before creating anything new.
The schema template
The SCHEMA.md template asks for lowercase hyphenated filenames, YAML frontmatter on every page and at least two outbound [[wikilinks]] per page.
It asks for 10 to 20 top-level tags, and every tag must be added to the taxonomy before anyone uses it.
It sets page thresholds, so a page is created when something appears in two or more sources or is central to one.
Pages are split once they pass about 200 lines and archived to _archive/ once they are fully superseded.
Frontmatter and quality signals
Every page carries title, created, updated, type, tags and sources in its frontmatter.
The optional confidence, contested and contradictions fields are the quality signals, and lint surfaces weak or disputed pages so they don't harden into accepted fact.
Pages that synthesise three or more sources get provenance markers such as ^[raw/articles/source-file.md] at the end of paragraphs.
The repo history shows these three features were ported into the skill from llm-wiki-compiler after launch, which is why my video from launch time doesn't cover them.
Ingest
Ingest captures a URL or PDF with web_extract, saves it under raw/ and adds frontmatter with the source URL, ingest date and a sha256 hash.
It discusses takeaways with you, checks for existing pages, writes or updates pages, updates the index and log, and reports every file it touched.
The docs say one source can update 5 to 15 pages, and the agent asks first if an ingest would touch 10 or more existing pages.
The hash means a re-ingest skips an unchanged source and flags drift when the source has changed.
Query
Query reads index.md, searches across files on bigger wikis, reads the relevant pages and answers with citations.
Substantial answers are filed back into queries/ or comparisons/, which is the part of the pattern I like most, because good analysis stops vanishing into chat history.
Lint
Lint is the strongest feature, with 13 documented steps.
| Lint check | What it catches |
|---|---|
| Orphans and broken links | It finds pages nobody links to and links that go nowhere. |
| Index completeness | It finds pages missing from index.md. |
| Frontmatter and tags | It finds missing fields and tags outside the taxonomy. |
| Stale content | It finds pages more than 90 days older than the newest source on the same entities. |
| Contradictions and quality | It surfaces contested pages and low-confidence claims. |
| Source drift | It recomputes raw file hashes and flags mismatches. |
| Size and log | It flags pages over 200 lines and rotates a log past 500 entries. |
Findings come back grouped by severity, with broken links first and style issues last, and the run is logged.
Bulk ingest and archiving
Bulk ingest reads every source first, checks existing pages in one pass, writes pages in one pass and logs the batch once.
Archiving moves a page to _archive/, removes it from the index and replaces links to it with plain text marked "(archived)".
Index scaling and log rotation
The index is sectioned into entities, concepts, comparisons and queries, with a one-line summary per page.
When a section passes 50 entries it gets split, and when the index passes 200 entries the agent creates _meta/topic-map.md.
The log uses the format ## [YYYY-MM-DD] action | subject and rotates to log-YYYY.md after 500 entries.
🔥 Want my full Hermes skill setup? Inside the AI Profit Boardroom, I've got a full Hermes Agent section with step-by-step tutorials and the installable Agent OS. You'll be learning alongside 3,400+ members with four coaching calls every week. → Get access here
Obsidian integration
The docs say the folder works as an Obsidian vault out of the box, with clickable wikilinks, Graph View and Dataview queries on the frontmatter.
They recommend setting Obsidian's attachment folder to raw/assets/, and they give a Dataview example query for filtering entity pages by tag.
If you also use the bundled obsidian skill, the docs say to point OBSIDIAN_VAULT_PATH at the same folder as your wiki.
Obsidian Headless for servers
For machines without a display, the skill documents obsidian-headless, installed with npm install -g obsidian-headless.
It needs Node.js 22 or newer and an Obsidian account with a Sync subscription, and ob sync --continuous keeps the vault syncing.
The docs include a systemd user service, so Hermes can write on a server while you read the same vault on your laptop or phone.
Updates and recovery
The docs say bundled skills are copied into ~/.hermes/skills/ on install and synced on every hermes update.
If you edit the skill, Hermes treats it as user-modified and stops updating it, which protects your changes.
The documented hermes skills reset llm-wiki command re-baselines it, and adding --restore brings back the pristine bundled version.
Hermes LLM Wiki vs RAG
| Question | Hermes LLM wiki | Typical RAG setup |
|---|---|---|
| When is knowledge processed? | It is compiled once at ingest and kept current. | It is retrieved from raw chunks at question time. |
| Does it accumulate? | It grows richer with every source and every filed answer. | It rediscovers the same knowledge on every question. |
| Can you read it? | You can browse every page in Obsidian or any editor. | The chunks sit in a store you rarely look at. |
| How are contradictions handled? | They are flagged in frontmatter and surfaced by lint. | They are left for the model to notice each time. |
| What infrastructure does it need? | It needs a folder of markdown files. | It usually needs an embedding pipeline and a vector store. |
Karpathy says an index file works surprisingly well at around 100 sources and hundreds of pages, so RAG still has a place for very large document piles.
Hermes Skill vs Karpathy's Original Gist
| Area | Karpathy's gist | The Hermes skill |
|---|---|---|
| Folder layout | He leaves it to you and your domain. | It fixes raw/, entities/, concepts/, comparisons/ and queries/. |
| Schema file | He suggests CLAUDE.md or AGENTS.md. |
It uses a dedicated SCHEMA.md with a full template. |
| Provenance | He doesn't prescribe a format. | It adds ^[raw/...] markers on multi-source pages. |
| Drift detection | He doesn't prescribe a method. | It stores a sha256 hash per raw source. |
| Lint | He lists the kinds of problems to look for. | It turns them into 13 ordered steps with a severity order. |
Hermes LLM Wiki vs llm-wiki-compiler
The docs point to llm-wiki-compiler, a Node.js CLI with the same Karpathy inspiration that batch-compiles a source folder into an Obsidian-compatible wiki.
They say it owns page generation and is tuned for small corpora, while the Hermes skill keeps the agent curating alongside you.
My view is simple, because you use the compiler for a one-off batch and the skill for a wiki you'll keep adding to.
The Limits
The skill is instructions, so a weaker model will write weaker pages and miss more cross-references.
Big ingests pause for approval when they would touch 10 or more existing pages, which is safe but slows unattended runs.
Raw sources sit on disk as plain files, so you need to decide where the folder lives and who can reach it.
The Obsidian Headless route needs a paid Obsidian Sync subscription, although the desktop app reads the folder for free.
If you want a second opinion on where an agent knowledge base fits your SEO business, you can book a free SEO strategy session with my team.
Who The Hermes LLM Wiki Is For
It is for anyone who researches the same subject over weeks or months and is tired of re-explaining it to an agent every session.
It suits researchers, content creators, agencies tracking competitors and teams that want an internal wiki nobody has to maintain by hand.
It is less useful for one-off questions, where a normal chat is faster.
If you're comparing it with other ways Hermes remembers things, my Hermes Agent memory rankings cover the alternatives, and my best Hermes Agent skills list shows what pairs well with it.
If you're bringing an existing Claude Code setup across first, my review of Hermes Import Agent covers that move.
🔥 Want help deciding which Hermes skills are worth your time? The AI Profit Boardroom has Hermes Agent training, the Agent OS and four weekly coaching calls where you can ask me which skills to switch on first. → Join the AI Profit Boardroom
Also On Our Network
- 🌐 How to set up the Hermes LLM wiki in seven steps
- 🌐 Six ways a business can put the Hermes LLM wiki to work
- 🌐 The Hermes LLM wiki quick start inside an Agent OS
- 🌐 Which models handle a Hermes LLM wiki best, per GoldieBench
FAQ: Hermes LLM Wiki Review
Is the Hermes LLM wiki any good?
I rate it 4.5 out of 5, because it is free, installed by default and documented in detail.
Its output quality depends on the model you connect to Hermes.
What is the difference between the Hermes LLM wiki and RAG?
RAG retrieves chunks from raw documents at question time and rediscovers knowledge on every query.
The Hermes LLM wiki compiles sources once into linked pages and keeps them current, so cross-references and contradictions are already in place.
How is the Hermes skill different from Karpathy's original LLM Wiki?
Karpathy's gist is intentionally abstract and leaves the folders and conventions to you.
The Hermes skill fixes a folder layout, a SCHEMA.md template, frontmatter fields, source hashing, provenance markers and a 13-step lint pass.
Does the Hermes LLM wiki work with Obsidian?
Yes, because the docs say the wiki folder works as an Obsidian vault out of the box.
They also document Obsidian Headless for servers without a display.
What does the Hermes LLM wiki cost?
The skill is free and MIT-licensed, and Hermes Agent is free and open source.
You pay only for the AI model you connect, unless you run a local model.
If you want to talk through how agents could take research work off your plate, book a free SEO strategy session and my team will help you plan it.
📺 Video notes + links to the tools 👉
🎥 Learn how I make these videos 👉
🆓 Get a FREE AI Course + Community + 1,000 AI Agents 👉
About Julian
I'm Julian Goldie, an SEO entrepreneur, author and founder of the AI Profit Boardroom, which has 3,400+ members.
I help business owners scale with AI agents, automation and SEO.
- I've built a 7-figure SEO and link-building agency, Goldie Agency, with 70+ people in Bangkok.
- I've grown a YouTube channel to 400,000+ subscribers.
- I wrote the book "Link Building Mastery", which covers nine link building strategies.
- I run GoldieBench, where I score AI models on real one-shot builds.
→ Get my best AI training inside the AI Profit Boardroom
Free, documented and built to compound, it earns a 4.5 out of 5 from me, and my verdict is to switch on the Hermes LLM wiki today.
