Smorster

Why LLM Wiki? 🧠 Future Of Knowledge For Agentic AI & Humans

Wanderloots · 11m 19s · Watch on YouTube · 15 sources

Decision Card

Effort: Weekend project — read Karpathy’s original gist (~15 min), install Obsidian plus the Web Clipper, create a separate “LLM vault,” and point one AI agent (e.g. Claude Code with a CLAUDE.md schema file) at it to compile clipped sources into wiki pages (~4–8 hours to a working prototype).

Honest take: The video is a concept explainer that defers all implementation to a promised future video, and it quietly conflates Obsidian’s untyped backlinks with real knowledge-graph triples — Obsidian edges have no named relationship (“caused”, “depends on”), so the demo graph lacks the very edge labels his own subject–relationship–object definition requires. The “AI maintains it automatically” framing also glosses over that Karpathy’s pattern needs you to author the schema/conventions file and trigger maintenance runs; nothing is automatic out of the box.

Concrete next steps:

  • Read Karpathy’s original LLM wiki gist to get the actual three-layer spec (raw sources / wiki / schema file): https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f (~15 min)
  • Prototype the pipeline: clip 5–10 articles with the Obsidian Web Clipper into a “sources” folder, then have an agent compile them into interlinked entity/concept pages in a sibling “wiki” folder (~2–4 hours)
  • If querying across many documents matters more than curation, evaluate Microsoft’s GraphRAG library directly instead: https://microsoft.github.io/graphrag/ (~1 hour to assess)
  • Skip if you don’t already have a note-taking habit or a steady stream of research sources — the whole value proposition is compounding over months, and an empty vault plus an agent produces nothing worth maintaining.

TL;DR

Callum (Wanderloots) explains knowledge graphs from first principles (nodes, edges, triples), shows how his 3-year Obsidian vault works as a personal knowledge graph, and argues for keeping a second, AI-maintained “LLM Wiki” vault — per Andrej Karpathy’s recent pattern — as shared structured memory across all your AI tools. The pitch: plain RAG fails when answers live between documents, so a persistent, interlinked markdown wiki that agents build and maintain outperforms per-tool siloed memory.

Key Points

  • The core proposal is two brains: a human vault for your own thinking and a separate LLM Wiki that AI agents build and maintain automatically, shared across every AI tool you use. 00:04
  • Knowledge graphs reduce to three primitives — node (thing/concept), edge (named relationship like “caused” or “depends on”), and triple (subject–relationship–object), the “atom” that compounds as you add knowledge. 01:07
  • Google’s knowledge panel is a live example: the sidebar for “Toronto Reference Library” is assembled by traversing Google’s Knowledge Graph, each fact a separate node. 01:50
  • Wikipedia works the same way — every internal link is an edge; a visualization of just 1.1% of Wikipedia dropped into Obsidian already looks like a dense galaxy of nodes. 02:24
  • In Obsidian, [[double-bracket]] links create nodes and edges as a side effect of writing — “the knowledge graph is just what happens when you’re specific about how you take notes.” 03:36
  • Standard RAG (embed, similarity-match, retrieve chunks) works for “what is X” but fails when the answer lives between documents — in the connections — which is where graph RAG comes in. 07:18
  • On large complex datasets, graph RAG significantly outperforms plain RAG and burns fewer tokens by following relationships instead of retrieving thousands of chunks. 08:00
  • Per-tool AI memory silos are the motivating pain: memory works inside one tool but is lost when you switch tools due to rate limits or performance — the LLM Wiki is the shared, tool-agnostic fix. 08:21
  • The LLM Wiki (popularized by an Andrej Karpathy article) has three layers: untouched raw sources (e.g. clipped via Obsidian Web Clipper), the agent-compiled interlinked wiki pages, and periodic agent maintenance for contradictions, stale info, and orphan pages. 09:56
  • He keeps human and LLM vaults firewalled so provenance stays clear: “That way I easily know what came from me and what came from AI.” 09:38

Notable Quotes

“A triple is the atom of a knowledge graph: subject, relationship, and object. That’s the entire model: two things and one connector.” 01:20

“I didn’t try to build the graph. I just wrote about the relationship between different concepts. The knowledge graph is just what happens when you’re specific about how you take notes.” 04:29

“The knowledge is compiled once and then kept current. It’s not rederived on every query.” 09:24

Verified Claims

  • Claim: The LLM Wiki concept gained popularity from an Andrej Karpathy article describing an LLM that “incrementally builds and maintains a persistent wiki — a structured interlinked collection of markdown files.” 08:48

    • Sources: Karpathy’s llm-wiki gist (published April 2026; the video’s quotation matches the gist verbatim, including the passage on updating entity pages and noting contradictions)
    • Verdict: Confirmed

  • Claim: Karpathy is formerly of OpenAI and coined the term “vibe coding.” 08:52

  • Claim: On larger complex datasets, graph RAG “significantly outperforms” plain RAG. 08:00

  • Claim: Google’s search sidebar (“knowledge graph panel”) is generated by traversing an entity-relationship map rather than by you clicking links. 01:50

  • Claim: RAG “converts your notes into numbers” (embeddings) and retrieves the most similar chunks, which works for simple lookups but struggles with relationship-spanning questions. 07:18

  • Claim: Someone dropped Wikipedia into Obsidian and visualized that 1.1% of it already forms a massive node-and-edge graph. 02:24

    • Sources: Obsidian Help: Graph view (mechanism — notes are nodes, links are edges — is real); the specific 1.1%-of-Wikipedia visualization could not be located in searches
    • Verdict: Inconclusive (mechanism confirmed, specific artifact unverified)

  • Claim (sponsor segment): The book on rebuilding civilization is “one of the top Kickstarter projects of all time.” 04:50

Tools, Papers & Standards Mentioned

Follow-up Questions

  1. How does an Obsidian-based LLM Wiki (untyped [[links]]) compare quantitatively to a true typed-triple store (e.g. GraphRAG’s entity graph or an RDF store) for multi-hop retrieval quality — do named edges actually matter for agent performance, or is link adjacency enough?
  2. What does the maintenance loop cost in practice — tokens and wall-clock per ingestion/consistency pass — and at what vault size does compile-once-and-maintain beat re-deriving via LazyGraphRAG-style query-time indexing?
  3. How should the human-vault/LLM-vault firewall handle bidirectional flow — when the agent’s wiki surfaces an insight worth promoting into the human vault, what provenance metadata prevents AI-generated content from silently contaminating “what came from me”?

Sources