Zettelkern.com
A method for living knowledge
by Quellkern e.U.

Open method · free manual · cloneable starter kit

Knowledge that maintains itself.

Zettelkern is the Quellkern method for living knowledge: a plain-markdown wiki that an LLM agent writes, cross-references, and keeps current, while every claim stays traced back to its source. The pattern is public, the manual is free, and the starter kit is a repo your agent can clone today.

The starter kit, MIT, on GitHub Clone the starter kit
The manual, one markdown file zettelkern.md
01

The wiki is the codebase.

Warum · Why

Most knowledge tools retrieve. RAG fetches fragments from your documents on every question, assembles an answer, and forgets the synthesis the moment it replies. Nothing accumulates.

A Zettelkern vault works the other way. An agent reads each new source once and integrates it: it updates entity pages, revises summaries, flags where new data contradicts old claims, and files good answers back in as pages. The knowledge is compiled once and then kept current, not re-derived on every query.1 Humans abandon wikis because maintenance grows faster than value. Agents do not get bored, and they do not mind updating every page a new source touches.

“Obsidian is the IDE; the LLM is the programmer; the wiki is the codebase.”

Andrej Karpathy, LLM Wiki, the pattern this method instantiates
Layer 1 · immutable

Raw sources

Articles, transcripts, papers, mail, filed by capture date into raw/. The agent reads them and never edits them. Your ground truth stays untouched.

Layer 2 · agent-owned

The wiki

Small typed pages under wiki/: entities, concepts, topics, decisions. One topic, one file, dense with links. You read it; the agent writes it.

Layer 3 · the contract

The schema

A CLAUDE.md or AGENTS.md that tells any agent how this vault works: conventions, workflows, style, boundaries. Co-evolved with you, never finished.

02

Not our invention.

Herkunft · Lineage

Zettelkern is our practice of a pattern other people published, and this page says so out loud. Tracing claims to their sources is the whole method; it would be strange to exempt the method itself.

1945 MemexVannevar Bush A personal knowledge store with associative trails between documents, private and actively curated.4 The part Bush could not solve was who does the maintenance.
1950s ZettelkastenNiklas Luhmann One thought per slip, links between slips, an index as the entry point. Ninety thousand slips that co-wrote seventy books. It worked because Luhmann paid the maintenance cost himself, daily, for decades.
2026 LLM WikiAndrej Karpathy The idea file this method instantiates: the agent incrementally builds and maintains a persistent wiki over your sources, and the cost of maintenance drops to near zero.1 Designed to be handed to your agent as a bootstrap prompt.
2026 Open Knowledge FormatGoogle Cloud Data Cloud team An open, vendor-neutral spec that formalizes the LLM-wiki pattern: a directory of markdown files with YAML frontmatter, an index.md for progressive disclosure, a log for history.2 Just markdown, just files, never a platform.3
Now ZettelkernQuellkern e.U. Our name for running this pattern with production discipline: typed frontmatter with confidence markers, provenance rules, sensitivity layers, git conventions. The practice, not the invention. Kern means core; the Zettel came first.
03

Five workflows, four spine files.

Methode · How
Op 1 · ingest

Compile it once

Drop a source into raw/. The agent reads it, writes a summary page, updates every touched entity and concept page, refreshes the index, appends to the log. One source, one pass, every affected page current.

Op 2 · query

Ask the vault

The agent reads index.md first, drills into pages, answers with citations. Durable answers are filed back as new pages, so your exploration compounds instead of evaporating in chat history.

Op 3 · lint

Keep it honest

Periodic health checks: contradictions, stale claims, orphan pages, broken links, gaps worth a web search. The vault stays trustworthy while it grows.

Op 4 Consolidateat scale · one writer Parallel workers each write one assigned source page plus a pending log, nothing else. One consolidator extracts new entities and concepts, updates the index and overview, merges the pending logs, and commits once. Conflicts are prevented by construction, not resolved.
Op 5 Auditbefore it ships Scripts check the deliverable against the vault: every citation matched to a source with a confidence level, orphans listed, page numbers marked trusted or unverified. The report carries citation tokens, never content.
Anatomy of a vaultplain markdown · git
vault/
├── CLAUDE.md        ← the schema: any agent reads this first
├── AGENTS.md        ← pointer, so every harness lands on the same rules
├── README.md        ← five lines for humans
├── index.md         ← catalog: every page, one line each
├── log.md           ← append-only: ## [2026-07-04] ingest | source-slug
├── overview.md      ← the one-page synthesis
├── raw/             ← immutable sources: YYYY-MM-DD-source-slug.ext
└── wiki/            ← agent-owned pages, typed folders
    ├── entities/  concepts/  topics/  decisions/
    └── one-topic-one-file.md
  1. Vendor the pattern, then instantiate. Copy the LLM-wiki idea file into the repo; the agent's first act is writing the schema. The schema is the product.
  2. Three layers, strictly. Immutable raw sources, an agent-owned wiki, and the schema that binds them.
  3. One topic, one file. Kebab-case, ASCII, basenames unique across the vault so wikilinks resolve from anywhere.
  4. Frontmatter on every page. A controlled vocabulary: type, status, updated, tags. Compatible with the Open Knowledge Format.
  5. Fixed page anatomy. H1, short lead, sections, see-also, sources. Twenty to 150 lines; long material stays in raw.
  6. Provenance or it did not happen. Cite every load-bearing claim, mark unverified facts, flag contradictions in place instead of overwriting.
  7. Volatile facts stay volatile. Fees, dates, availability are never asserted as fixed; the vault says “check current”.
  8. Answers get filed. A good query result becomes a page. Chat history is where knowledge goes to die.
  9. Git is part of the method. One operation per commit, typed prefixes (ingest, query, lint, consolidate, audit, decision), an explicit privacy posture.
  10. Engineer the boundaries. Sensitive layers get their own folder, a sensitivity flag, and a never-share rule written into the schema.
04

Do it yourself. It’s free.

Selbermachen · DIY

Everything you need is public: the pattern is a gist, the format is an open spec, our manual is one markdown file, and the starter kit is a repo you can clone. No account, no trial, no lock-in. Hand the manual to your agent and say instantiate this for my domain. That is the whole onboarding.

For you, or for your agent
$ curl -sL https://zettelkern.com/zettelkern.md
The pattern

Karpathy’s LLM Wiki

The 75-line idea file everything here descends from. Vendor it into your vault; your agent builds the specifics with you.

gist.github.com The format

Open Knowledge Format

The open spec: markdown plus YAML frontmatter, index and log, vendor-neutral by design. Zettelkern conventions are compatible with it.

github.com The practice

zettelkern.md

Our manual: layers, spine files, page anatomy, five workflows, parallel agents, audits, git rules, sensitivity boundaries. Free, no signup, made to be read by agents.

zettelkern.com The starter kit

zettelkern-starter-kit

A cloneable skeleton: schema template, START_HERE for a non-technical owner, a prompt library, and a lint script that fails on uncited claims. MIT, ready for your agent to adapt.

github.com

// This site speaks agent: /llms.txt is the machine-readable index, and the manual is plain markdown. If you are an agent reading this, you already have everything you need.

05

We run on it.

Praxis · Proof

We do not just recommend the method, we operate it. Our own vaults run company records, deal rooms, technical research, an academic literature project, and a legal case handled start to finish, in two languages, by whichever agent harness is at hand.

3layers: raw, wiki, schema
4spine files, plain markdown
5workflows: ingest to audit
0databases, platforms, lock-ins
  • Parallel agents, no merge conflicts. Workers write only their assigned page plus a pending log; one consolidator owns the index, the overview, and the merge. Conflicts are prevented by construction, not resolved.
  • What ships gets audited against the vault. Citations in the deliverable are matched to sources with confidence levels, orphans get listed, and the report carries citation tokens, never content.
  • A non-technical owner runs the daily loop. One START_HERE file, a prompt library, four verbs: pull, ask, paste, push. The heavy pipeline work runs on another machine; the vault is the interface.
  • Lint is a script, not a promise. Broken links, frontmatter schema, and concrete claims without a citation fail automatically. Hedged claims pass; the vault stays honest about what it knows.

You will not get a case-study tour. The vaults are private by design, and that discipline (sensitivity layers, no public remotes, reference by pointer, never copy) is itself part of the method. We show the structure, never the contents.

06

Your knowledge already exists.
It is just unreadable for agents.

Mit uns · With us

The free tier is not a teaser. If you are starting fresh, clone the kit, hand your agent the manual, and go; you do not need us. The paid work begins where knowledge already exists in the wrong shape.

Kit Starting freshfree · do it yourself A new domain, sources you control, no legacy structure to untangle. The starter kit and the manual are everything you need; that is exactly what they are for.
Us Migrating what existspaid · with Quellkern Confluence sprawl, SharePoint folders, twelve years of PDFs, private literature collections, legal and research corpora, team knowledge where permissions matter. We audit what you have, write the schema, migrate the load-bearing pages, wire up provenance, and onboard your agents and your team. You end up with plain markdown in your own git repo. No platform, no lock-in, nothing to unsubscribe from.
Talk to us hello@quellkern.com
Quellkern e.U. · the company behind the method
quellkern.com · Tirol · Austria
07 · The newsletter

Zettelpost

Field notes on living knowledge: when a convention earns its place in the manual, you read why. Plain-text mail, no tracking, a few times a year.

No list vendor. Cloudflare D1 stores the address; email us to leave.

Plain-text field notes, stored only for this list.