Karpathy's “LLM Wiki,” Made Effortless: An AI-Maintained Knowledge Base Without the Setup

Key points

• The LLM Wiki is a pattern where, instead of retrieving from raw documents at query time, an LLM incrementally builds and maintains a persistent, interlinked Markdown wiki.
• It is powerful, but you have to assemble and keep running a local stack: Obsidian, Claude Code, Git, and a hand-designed schema.
• Jade Note delivers that same “AI grows and organizes your knowledge” workflow with zero setup — just sign up and write.

The “LLM Wiki,” shared by Andrej Karpathy, is a personal-knowledge pattern quietly gaining attention. In one line: instead of retrieving from raw documents on every query (RAG), the LLM incrementally builds and maintains a persistent, interlinked Markdown wiki. Because knowledge is compiled once and kept current, the cross-references are already there, and any summary reflects everything you’ve read — from the start.

What the LLM Wiki is

The LLM Wiki rests on three layers.

• Raw Sources — immutable documents: articles, papers, PDFs, images. The LLM only reads them; it never edits them. This is the ground truth.
• The Wiki — LLM-generated, LLM-maintained Markdown pages: summaries, entity pages, concept pages, cross-links, and an overview that ties everything together. By reading index.md first (a one-line catalog of every page), the agent finds the right page without embedding-based retrieval. log.md is a dated, append-only record you can query with ordinary tools like grep.
• The Schema — a config document such as CLAUDE.md. It defines the wiki’s structure, naming conventions, and the workflows for ingestion, querying, and maintenance.

The workflow runs as an Ingest → Query → Lint loop.

• Ingest — when you add a new source, the LLM reads it, updates the relevant entity and concept pages across the wiki, appends to the log, and refreshes cross-references. A single source can ripple through 10–15 pages.
• Query — you ask a question; the LLM consults index.md, reads the relevant pages, and answers with citations. Useful explorations get written back as new pages, so knowledge compounds.
• Lint — periodically, you ask the LLM to health-check the wiki for contradictions, stale claims, orphan pages, missing cross-references, and gaps.

Karpathy’s sharp insight is that the real burden of a knowledge base isn’t reading or thinking — it’s the bookkeeping: updating cross-references, keeping summaries current, reconciling contradictions. Humans abandon wikis precisely because that maintenance cost outpaces the value. “The LLM doesn’t get bored, doesn’t forget to update a cross-reference, and can touch 15 files in one pass.” It’s also the long-awaited answer to the one part of Vannevar Bush’s 1945 Memex — a curated personal knowledge store with associative trails — that he couldn’t solve: who does the maintenance. The LLM does.

The catch — the weight of setup and upkeep

The idea is beautiful. The problem is that the road to actually running it is long.

• Install Obsidian and design the structure of your Vault
• Set up Claude Code and get comfortable driving an agent from the terminal
• Configure Git for version control (so you can safely let the AI write)
• Build a pipeline to turn articles into Markdown with a Web Clipper or similar
• And, crucially, design and keep maintaining the CLAUDE.md schema and your index.md / log.md conventions yourself

And this is a local-first workflow. Syncing across devices, or reaching it from “the ChatGPT I always use on my phone,” is something you have to engineer yourself. In short, the LLM Wiki is powerful, but its benefits largely accrue to engineers who can assemble the stack and keep it running.

Jade Note: the same workflow, zero setup

The point isn’t that the LLM Wiki is wrong — it’s that the idea is right, and it can be far more effortless.

Jade Note ships the very pieces you’d otherwise hand-assemble as built-in features from day one. Sign up, split things into categories, and write Markdown. The AI takes it from there — organizing, cross-linking, and maintaining.

LLM Wiki building block	The equivalent in Jade Note
The Wiki & cross-links	Category Wikis + AI-suggested links and typed links (related / parent / child / continues / references)
Querying via index.md	AI-indexed semantic search & keyword search
Lint (contradiction / staleness checks)	Note freshness checks & automatic Category Wiki updates
Version control with Git	Built-in versioning, change preview, and rollback
Designing CLAUDE.md / a schema	No design required — just split into categories and write
Standing up Obsidian + Claude Code locally	MCP-native — the Claude / ChatGPT you already use reads and writes in place

That last row is the decisive one. With an LLM Wiki, you open a terminal and fire up Claude Code to use the AI. With Jade Note there’s no new tool to learn, because the Claude or ChatGPT you already use all day reads and writes your notes directly through MCP (the Model Context Protocol).

• Claude or ChatGPT runs semantic search over your notes right inside the conversation (that’s Query)
• The AI saves and appends facts and decisions from the conversation into your notes on its own (that’s Ingest)
• The AI takes on the upkeep — link suggestions, Category Wiki updates, freshness checks (that’s Lint)

In other words, think of Jade Note as a cloud LLM Wiki that runs the Ingest → Query → Lint loop for you, without you assembling any of the stack.

Which should you choose?

If you want everything local, designing the files and the schema by hand for full control — and you enjoy that kind of tinkering — the LLM Wiki is a fantastic playground. Following Karpathy’s gist and the community implementations it spawned is well worth your time.

But if you’d rather not open yet another app or terminal, and you want the ChatGPT or Claude you already use to simply remember and organize things for you, try Jade Note. The more the LLM Wiki idea resonates with you, the more Jade Note’s effortlessness will land. Just write — the AI grows your knowledge.

Get started free

Source: Andrej Karpathy — LLM Wiki gist