Breaking

Karpathy's 'LLM Wiki' Turns Obsidian and Claude Into a Persistent Knowledge Base

July 1, 2026 at 20:15 EDT

AI Agents
Software Dev & Coding
Foundation Models

Andrej Karpathy (former OpenAI/Tesla AI lead) has drawn attention with an "LLM Wiki" pattern published in April 2026 that pairs Obsidian with Claude to have an LLM incrementally build and maintain a persistent, interlinked Markdown wiki from raw sources. One writer reportedly fed in their entire body of essays and generated 56 interlinked pages in one hour.

April 2026 · The Karpathy Pattern

The "LLM Wiki": knowledge that compiles once and compounds forever

A GitHub idea-file pairing Obsidian as the "IDE" and Claude as the "programmer" is spreading fast. One user fed in all of their essays and got 56 interlinked pages in about an hour — unlike RAG tools that "rediscover" everything with every query.

95%

Claude token usage cut by converting 383 files into a compact wiki

Interlinked pages built from one user's essays in ~1 hour

383

Raw files compiled into a single persistent knowledge base

Token cost: compile once vs. re-read every time

383 files fed to Claude — before and after the raw/wiki two-layer structure

100%

Conventional RAG
(re-read per query)

LLM Wiki
(compiled once)

Compiling knowledge into a persistent wiki cut token use by 95%.

The three-action loop

Ingest

Bring sources in and integrate them into the wiki

→

Query

Answer from the wiki, then feed the answer back in

→

Lint

Detect contradictions and orphaned pages

Obsidian = IDE · LLM = programmer · Markdown wiki = codebase — knowledge stacks up instead of vanishing.

What users praise

Surprising links across three years of notes
Automated link maintenance, less effort
Graph view for visual grasp
Local Markdown — git history, no lock-in

Where it strains

Hallucinations create faulty links — needs human review
Long PDFs/books are hard (go chapter-by-chapter)
Initial schema (CLAUDE.md) setup is fiddly
Search performance questions at scale

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.