Here's what happened.

Everyone's talking about Karpathy's LLM Wiki. 5,000 GitHub stars in three days, trended on X, dozens of open source repos. And there's zero code in it. Just a markdown file with an idea.

So why did people lose their minds?

Because it nails something everyone's been feeling. Sure, LLMs have memory now. ChatGPT remembers your name, Claude remembers your preferences. But that's just sticky notes. Nobody is actually *building* anything with all the research we do every day. It just... evaporates.

Karpathy's idea: stop using LLMs to answer questions. Start using them to build knowledge. Have the LLM maintain a persistent wiki. Structured pages, cross-references, evolving synthesis. Every source you add makes the whole thing smarter.

As he put it: "The tedious part of maintaining a knowledge base is not the reading or the thinking — it's the bookkeeping. LLMs don't get bored, don't forget to update a cross-reference, and can touch 15 files in one pass."

Before, LLMs had to start from scratch with every new query. Now, they get a head start.

What we built

Karpathy designed the wiki layer. We built the data layer that feeds it. Our integration with Obsidian simplifies and accelerates the web search process needed to feed the wiki.

We just shipped this as the memory architecture for our open-source Nimble Web Search Skills. You run a skill (i.e. competitor intel, meeting prep, company research) and it doesn't just search the live web and generate reports, it digests and organizes the results straight into a structured, linked wiki

Importantly, these wikis aren't just lists of pages. The content is intelligently mapped and continuously updated to make it easy for LLMs to understand the relationship between the various information.

And each run makes the next one smarter. Yesterday's research enriches today's queries. Over time you end up with your own Private Web Index.

The stack is simple:

Nimble (search + extract)    →  live web data

         ↓

LLM (compile + cross-link)   →  wiki pages

         ↓

Obsidian (navigate + graph)  → human layer

• Nimble searches the live web, such as news, social, financial data, company pages, and forums, and extracts clean markdown. 
• The LLM compiles it into cross-linked wiki pages with dates, sources, and entity relationships. 
• Obsidian renders the whole thing with its graph view, you see every connection, every entity, every trail.

All local. All yours. Plain markdown files.

‍

We built two indexes from scratch to prove it works

Tech/AI: The Karpathy LLM Wiki Ecosystem

I wanted to understand the full picture: Obsidian, Karpathy, and what happened when the gist went viral.

What Nimble found:

• Obsidian: bootstrapped, $25M ARR, $350M valuation, 9 employees (3 engineers)
• Karpathy's path: OpenAI co-founder → Tesla AI director → Eureka Labs → LLM Wiki
• The viral moment: 5K stars in 48hrs, X trending, dozens of repos, Steph Ango responded
• Enterprise adoption: LinkedIn built a coding agent knowledge base (20% adoption lift), Shopify CEO ran the AutoResearch pattern overnight

Result: 101 cross-linked wiki pages: 35 companies, 16 people, 14 events, 13 concepts, 7 projects.

Healthcare: The GLP-1 Drugs Market

Completely different domain. Same approach. It all starts with one prompt in Claude Code:

/nimble:market-finder GLP-1 Drugs Market

What Nimble found:

• $52-94B market (2025), projected $132-185B by 2033
• Eli Lilly became the first $1T healthcare company on tirzepatide momentum
• Novo Nordisk CEO replaced after CagriSema missed its target
• Pipeline: retatrutide (triple agonist, ~24% weight loss), orforglipron (first unrestricted oral), amycretin (next-gen)
• FDA cracking down on compounded semaglutide — Hims & Hers stock collapsed
• Medicare Part D obesity coverage starting July 2026 — 40M new beneficiaries

Result: 82 cross-linked wiki pages: 18 companies, 10 drugs, 10 clinical trials, 7 people, 6 markets, 5 regulatory events.

Watch the graph grow, each node is a wiki page Nimble built automatically:

Two completely different verticals. Same pattern. That's the point, the wiki architecture works for anything. You just need the data flowing in.

How to try it

Install the Nimble Agent Skills plugin (it handles the CLI automatically):

claude plugin marketplace add Nimbleway/agent-skills && \

claude plugin install nimble@nimble-plugin-marketplace

Then open Claude Code and say:

Run company-deep-dive on [any company]

The skill will search the web, extract the best sources, compile cross-linked wiki pages, and save everything to `~/.nimble/memory/` as plain markdown.

Open that folder in Obsidian. Hit `Cmd+G` for the graph view. That's your Private Web Index.

Run more skills. Each one builds on what's already there:

Run competitor-intel on [company]

Run meeting-prep for [person]

The graph gets denser. The connections multiply. The index compounds.

What others are doing with this pattern

Since Karpathy's post:

- Lex Fridman has the LLM generate a temporary mini-knowledge-base that he loads into voice mode for 7-10 mile runs

- Tobias Lütke (Shopify CEO) ran Karpathy's AutoResearch pattern — 700 experiments over 2 days, made Liquid template engine 53% faster

- LinkedIn built CAPT — a contextual knowledge base powering AI-assisted development for 1,000+ engineers, cutting initial triage time by ~70%

- A welder on X built a domain knowledge base: "Just markdown, FTS5, and grep. Every bug fix gets indexed. The knowledge compounds."

I love that last one. A welder and a Stanford AI professor using the same pattern. That tells you something.

Where this goes

Right now the wiki lives on your laptop. That's intentional — local, private, yours. But the pattern opens up some interesting doors:

• Team wikis — multiple agents feeding one shared knowledge base, with human review in the loop
• Continuous monitoring — skills that run on a schedule, updating entity pages as the web changes
• Fine-tuning on your own wiki — knowledge baked into model weights, not just context

We're exploring all of these. The Nimble Web Search agent-skills repo is open source — if you build something interesting with it, I'd love to see it.

Resources:

• Agent Skills Repo
• Karpathy's LLM Wiki Gist
• Obsidian (Free for personal use)
• Nimble CLI Docs

‍

‍