PDFSQLMDTS
Every question starts from fragments.
Retrieve chunks. Reconstruct context. Generate an answer. Discard the useful synthesis.
Interactive field guide Agent-native knowledge
Your AI shouldn’t rediscover what you already know.
Files can be retrieved. Knowledge should accumulate. Build a living wiki, watch OpenWiki maintain it, then move it anywhere with OKF.
3real-world missions
0uploads or API calls
1durable mental model
KNOWLEDGE COMPILERSIMULATION / LIVE
LLMWIKIpersistent artifact
TSauth.ts
MDnotes.md
PDFpaper.pdf
SQLmetric.sql
01 / The shift The answer is not the asset.
From retrieval to accumulation
The answer is not the asset.
The wiki is.
Every answer can improve the artifact.
Sources become concepts, relationships, contradictions, guides, and logs that remain available tomorrow.
Guided simulation All sample inputs and outputs are precomputed. No files leave your browser.
02 / The lab
Direct manipulation, not another explainer
The knowledge workbench
Choose a world. Move its scattered context into a persistent artifact. Then inspect what changed.
Mission 01 · Codebase memory
Onboard without the archaeology.
The answer is scattered across routes, middleware, configuration, tests, and unwritten conventions.
The expensive question“Where is authentication handled?”
Waiting for sources
Uncompiled sources
5 filesScattered context
router.post('/session', validateCredentials, createSession);const session = await sessions.verify(request.cookies.get('sid'));export const sessionTtl = 60 * 60 * 24 * 7;expect(response.headers.get('set-cookie')).toContain('HttpOnly');Rotate signing keys through SESSION_KEY_CURRENT and SESSION_KEY_PREVIOUS.Compile
0 of 5 sources compiled
Persistent artifact
WaitingLiving wiki
Move knowledge hereDrag a sample file or use its arrow button.
03 / Relationships
A folder becomes a model
Knowledge is more than files.
Each edge makes a claim visible: what implements, governs, qualifies, or supports the answer.
- Routes + middleware implements Authentication
- Policy + tests governs Session lifecycle
- Authentication creates Session lifecycle
- Authentication grounds Contributor answer
- Session lifecycle qualifies Contributor answer
04 / The difference
Retrieval versus accumulation
Ask once. Keep the work.
RAG retrieves fragments for this answer. An LLM Wiki leaves behind a maintained artifact for the next answer.
Choose a question. Watch two knowledge systems take different paths to an answer.
05 / Operationalize it
From pattern to workflow
OpenWiki does the maintenance work.
It scans a repository or configured sources, synthesizes local wiki pages, connects coding agents, and records what changed.
Run a deterministic OpenWiki simulation. No repository or API key is used.
Guided simulation · precomputed output
06 / Make it portable
Format, not platform
OKF gives knowledge a shared shape.
Markdown stays readable. YAML makes concepts queryable. Reserved index and log files make the bundle navigable and auditable.
acme-auth-knowledgeOKF v0.1 · portable bundle
code agent can read the same bundle—without a proprietary SDK.
What this file does
Repository knowledge index
Progressive entry point for agents and humans.
07 / The whole system
Three layers, one durable loop
Remember this relationship.
Idea
LLM WikiCompile source material into a living, interlinked knowledge artifact.
Tool
OpenWikiGenerate, connect, and maintain agent-oriented wiki pages.
Format
OKFExchange the resulting knowledge as portable Markdown and metadata.
Your next working session
Give every contributor the map you wish you had.
Faster onboarding, cheaper agent searches, and documentation that changes with the code.
- 01Run OpenWiki in code mode at the repository root.
- 02Review the first architecture pages and add scope notes.
- 03Update the wiki in CI when code changes.
08 / Plain language
Keep the vocabulary precise
A five-term glossary.
- Source
- The original material: code, papers, notes, tables, decisions, or conversations.
- LLM Wiki
- A persistent, interlinked knowledge artifact that an LLM helps compile and maintain.
- RAG
- A pattern that retrieves source fragments at question time to ground a generated answer.
- OpenWiki
- LangChain’s open-source CLI for writing and maintaining agent wikis from codebases or configured sources.
- OKF
- A vendor-neutral Markdown and YAML format for portable, agent- and human-readable knowledge.
09 / Questions
Before you try it
Clear answers, no mystique.
01Is LLM Wiki a product?+
No. LLM Wiki is a pattern proposed by Andrej Karpathy: an LLM compiles sources into a persistent, interlinked wiki that can improve as it is used.
02How is an LLM Wiki different from RAG?+
RAG usually retrieves source fragments again for each question. An LLM Wiki leaves behind maintained synthesis pages, links, questions, and logs so useful reasoning becomes reusable knowledge.
03What does OpenWiki do?+
OpenWiki is LangChain’s open-source CLI for writing and maintaining local agent wikis from codebases or configured personal knowledge sources.
04What is OKF?+
The Open Knowledge Format is a vendor-neutral specification for portable knowledge bundles made of Markdown files, YAML frontmatter, and simple index and log conventions.
05Does this page upload my files or call an LLM?+
No. This is a deterministic guided simulation using precomputed sample sources and outputs. Nothing is uploaded and no model API is called.
10 / Primary sources
Read the originals
Source-verified, not trend-summarized.
01 · IdeaAndrej Karpathy — LLM WikiOriginal gist · April 4, 2026 02 · ToolLangChain — OpenWikiSource, install, modes, and operations 03 · AdoptionOpenWiki 0.2 adds OKFLangChain · July 16, 2026 04 · FormatGoogle Cloud — Introducing OKFAnnouncement · June 12, 2026 05 · SpecificationOpen Knowledge Format v0.1Apache-2.0 specification