Earlier this week, Andrej Karpathy — the former head of AI at Tesla and one of the most respected voices in machine learning — published a gist on GitHub that quietly reshapes how we think about knowledge management. His proposal: stop treating LLMs as search engines over your documents. Instead, let them build and maintain a living wiki that grows smarter with every source you feed it.
The idea is deceptively simple. But the applications are far-reaching — especially for professionals who deal with complex, evolving information every day.
The core idea in 30 seconds
Traditional RAG (Retrieval-Augmented Generation) works like this: you ask a question, the system searches your documents, retrieves relevant chunks, and generates an answer. It works, but every query starts from scratch. There’s no memory, no synthesis, no accumulated understanding.
Karpathy’s LLM Wiki flips this. When you add a new source — an article, a meeting transcript, a research paper — the LLM doesn’t just index it. It reads it, extracts the key insights, updates existing wiki pages, adds cross-references, flags contradictions with earlier sources, and logs what changed. The result is a persistent, compounding knowledge base that reflects everything you’ve ever fed it.
Three layers make it work:
- Raw Sources — your original documents, untouched
- The Wiki — markdown pages created and maintained entirely by the LLM
- The Schema — a configuration document (like a CLAUDE.md file) that tells the LLM how to behave
As Karpathy puts it: “The wiki is a persistent, compounding artifact. The cross-references are already there. The contradictions have already been flagged. The synthesis already reflects everything you’ve read.”
Five applications that make this practical
The pattern is deliberately abstract — Karpathy designed it to be adapted, not copied. Here are five concrete ways professionals can put it to work.
1. A research wiki that actually remembers what you’ve read
If you’ve ever spent weeks researching a topic — reading papers, bookmarking articles, highlighting passages — only to realise months later that you can’t find that one critical insight, this pattern is for you.
How it works: Create a wiki for your research domain. Every time you find a relevant source, drop it in the raw sources folder. The LLM reads it, creates a summary page, and — crucially — updates every related concept and entity page in your wiki. A new paper on transformer architectures doesn’t just get its own page. It updates your “Attention Mechanisms” page, adds a comparison to your “Architecture Comparison” page, and flags where its findings contradict an older source.
Why it matters: After 50 sources, you don’t have 50 isolated summaries. You have an interconnected knowledge graph where every insight is contextualised against everything else you’ve read. The LLM handles the cross-referencing that humans inevitably abandon.
2. Client intelligence that compounds over time
Consultants, account managers, and agencies deal with a familiar problem: institutional knowledge about clients lives in scattered emails, call notes, and people’s heads. When someone leaves, the knowledge walks out with them.
How it works: Set up a wiki per client (or per project). Feed it meeting transcripts, email threads, project briefs, and feedback notes. The LLM maintains pages for key stakeholders, ongoing decisions, open questions, and historical context. When you prepare for a meeting, ask the wiki — it already has a synthesis of everything you’ve discussed.
Why it matters: You stop relying on human memory for continuity. A new team member can read the client wiki and understand in an hour what took the previous consultant months to accumulate. The maintenance happens automatically — every new source updates the whole picture.
3. A personal learning wiki that grows with you
Online courses, books, podcasts, conference talks — most professionals consume a staggering amount of content. Almost none of it sticks in a structured way.
How it works: Create a wiki for a skill you’re developing. When you finish a book chapter, a course module, or a talk, feed the notes into your wiki. The LLM files concepts into the right pages, links related ideas across different sources, and tracks your evolving understanding. Ask it to generate a “what I know about X” synthesis at any point — it draws from everything, not just the last thing you read.
Why it matters: Learning compounds when you connect new ideas to existing knowledge. The wiki makes those connections explicit and persistent. It’s the difference between a pile of highlights and a structured understanding.
4. Competitive intelligence that stays current
Tracking competitors, market trends, or regulatory changes is a classic knowledge management problem. The information changes constantly, and today’s analysis is outdated by next quarter.
How it works: Build a wiki with pages for each competitor, market segment, and key trend. Feed it analyst reports, news articles, product announcements, and regulatory updates. The LLM maintains a living competitive landscape — updating competitor pages when new information arrives, flagging shifts in positioning, and noting when a trend page hasn’t been updated in a while (the “lint” workflow Karpathy describes).
Why it matters: Instead of starting a competitive analysis from scratch every quarter, you have one that’s continuously maintained. The lint workflow catches staleness — orphan pages, missing cross-references, claims that newer sources contradict — so your intelligence stays trustworthy.
5. Team knowledge that doesn’t decay
Every team has a wiki problem. Someone sets up Notion or Confluence with the best intentions. Six months later, half the pages are outdated and nobody trusts the content anymore.
How it works: Configure the LLM Wiki to ingest your team’s natural outputs — Slack threads, meeting transcripts, decision logs, post-mortems. The LLM distills these into structured wiki pages: process documentation, decision records, onboarding guides, FAQ pages. When a new decision contradicts an old one, the wiki flags it and asks for resolution.
Why it matters: Traditional wikis decay because maintenance is manual and boring. LLMs don’t get bored. The cross-referencing, consistency checking, and page updates that kill human-maintained wikis are exactly what LLMs do well. The human role shifts from maintenance to curation — deciding what matters, not keeping pages up to date.
What makes this different from “just using ChatGPT”
You might wonder: can’t you already do this by dumping documents into a ChatGPT conversation? Not quite. Three things set the LLM Wiki pattern apart:
Persistence. Chat conversations are ephemeral. The wiki is a file system. It survives across sessions, can be version-controlled with git, and doesn’t depend on any single provider.
Compounding. Each new source doesn’t just get processed — it enriches every related page. Source number 50 benefits from the context of the previous 49. In a chat, you’d have to re-upload everything or hope the context window is large enough.
Structure. The schema layer means the LLM behaves consistently. It follows your conventions for page format, tagging, cross-referencing, and quality. It’s not a free-form conversation — it’s a disciplined knowledge system.
Getting started
Karpathy intentionally kept the proposal abstract. There’s no repository to clone, no app to install. The pattern is designed to be instantiated with whatever tools you prefer.
A practical starting point:
- Pick a domain — one research topic, one client, one learning goal
- Create a folder structure —
sources/,wiki/, and a schema file (e.g.,CLAUDE.md) - Write your schema — define how pages should be structured, how ingest should work, what conventions to follow
- Start feeding sources — let the LLM build the initial wiki pages
- Iterate — refine the schema as you learn what works
Tools like Obsidian work well as a frontend — you get markdown editing, graph visualisation, and a plugin ecosystem. But a plain folder of markdown files works too.
The key insight is that this isn’t a product. It’s a workflow pattern that leverages what LLMs are genuinely good at — reading, summarising, cross-referencing, and maintaining consistency — while keeping humans in the role they’re best at: deciding what matters.
What’s next?
- Want to understand the tools that make this possible? Read Claude Cowork: a practical guide for beginners
- Ready to build something more structured? Try Build your own AI skill in 15 minutes
- Curious how AI fits into your daily workflow? See Five daily tasks you can automate with AI
Source: Andrej Karpathy — LLM Wiki (GitHub Gist, April 2026)