Fix Karpathy's LLM Wiki with a Knowledge Graph | Claude Code + Obsidian + InfraNodus

Nodus Labs — Fix Karpathy’s LLM Wiki with a Knowledge Graph (Claude Code + Obsidian + InfraNodus)

A 26-minute YouTube tutorial by Nodus Labs (creators of InfraNodus, a knowledge-graph visualisation tool the speaker says they’ve been building “for the last 10 years”). Published 11 April 2026 — one week after Karpathy’s LLM Wiki gist (4 April 2026) — making this the wiki’s earliest third-party “fix” / extension proposal for the LLM Wiki pattern, and one of three early-April / mid-April LLM-Wiki engagements already in the wiki (OceanBase ex-brain / Raju explainer / this video).

The presenter is the InfraNodus founder (named in the description as Dmitry Paranyushkin — the speaker self-identifies as having worked on InfraNodus for ten years, consistent with that attribution; the wiki defers a named-entity promotion to next ingest for verification). Channel-style: hands-on tutorial with screen-recording demo. The video positions itself as an immediate response to a perceived limitation in Karpathy’s pattern.

TL;DR

• The diagnosis (the load-bearing claim). Karpathy’s LLM Wiki is a “really good framework” but “not aware of itself.” When a user queries the wiki, the LLM scans the structure, extracts concepts, and “produces the most probable outcome — because that’s what LLMs are trained to do.” Result: the wiki has structure, but querying it still yields generic outputs because the LLM is using flat-prompt next-token logic to navigate it. “Even though you have the structure, you still need something to navigate through it.”
• The proposed fix — knowledge-graph augmentation, three integration depths.
  1. External tool — open InfraNodus’s web app on the wiki content; visualise clusters and gaps; copy back insights to Claude.
  2. MCP server — connect InfraNodus to Claude Code via MCP; generate_knowledge_graph and gap-analysis tools become directly callable from inside the LLM session. “You can also use the InfraNodus MCP server inside Claude Code itself, which is basically an API that connects this whole content to InfraNodus.”
  3. KG-integrated wiki (the deepest layer) — the Nodus Labs skill bakes knowledge-graph generation into the LLM Wiki workflow itself. After ingest, the wiki contains both concepts/ markdown files and an infranodus/ folder holding ontology graphs generated on every interaction. “It’s some kind of living memory of what’s happening in the system — every time you come back to it, Claude (or whatever system you use) will have access to this system of rules of how the concepts connect in your wiki.”
• The novel mechanism — KG-as-attention-direction, not just KG-as-retrieval-substrate. The gap-analysis workflow: (1) compute clusters from concept-page co-occurrence; (2) find pairs of clusters with low betweenness (the “gap”); (3) copy the gap prompt (a structured representation of two clusters and their disconnection) back to Claude; (4) ask Claude to generate insights that bridge the two clusters using the underlying source documents. “I point the LLM’s attention to the gap that exists. I provide the underlying structure. I give it some context, and then I say okay, use the documents in the concepts folder in order to generate an interesting idea that would connect these ideas together.” The KG isn’t substituting for retrieval — it’s steering attention to where the wiki is missing connections.
• The opinionated stack (worked example demonstrated in screen-record):
  • Obsidian — viewer / human-navigation layer. Includes the InfraNodus Obsidian plug-in for graph view alongside the wiki.
  • Claude Code / Cursor — LLM interaction layer.
  • InfraNodus VS Code extension + InfraNodus Obsidian plug-in + InfraNodus MCP server — three install paths into the same KG primitives.
  • The Nodus Labs skill-llm-wiki Claude skill — published at github.com/infranodus/skills/blob/master/skill-llm-wiki/SKILL.md — the first wiki source on a third-party-shipped Claude skill that operationalises the LLM Wiki pattern.
• “Living memory” framing. Every Claude session generates new InfraNodus ontology graphs that get saved alongside the wiki. The wiki’s concepts/ are what is known; the infranodus/ graphs are how the concepts relate, captured at successive points in time. The graph layer is treated as compounding state in its own right — not just a derived view. This is structurally parallel to Osmani’s “harness is a living system” claim applied at the wiki-content layer rather than the agent-runtime layer.
• The gap-analysis demo (Chapter 13–14, 18:00–22:00). The speaker walks through a real wiki on a research topic (“financial flows” + “regression analysis” identified as two clusters with a gap between them). InfraNodus emits a structured gap-prompt that names the two clusters, the candidate bridge questions, and the most-dense source extracts from each cluster. Claude reads this gap-prompt + the underlying concept files and generates a bridging analysis. The output is markedly less generic than a flat-prompt query of the same wiki.
• The MCP-server-without-graph-UI option (Chapter 16, 23:15-end). If the user doesn’t want to open the graph UI, they can simply tell Claude “run gap analysis on the wiki” and Claude uses the InfraNodus MCP gap_analysis tool to perform the same workflow — read all concept files, identify clusters, find gaps, generate bridging questions, save the ontology graph back to infranodus/. First wiki source on KG-gap-analysis as a callable agent primitive, not just a human-visualisation tool.
• Closing recommendation to the channel’s audience. “You can use the graph if you want, or you can also use the MCP server inside Claude Code itself. … If you feel uncomfortable using the graph interface, you can just delegate all this responsibility to your LLM.” Pragmatic — both interfaces are valid; the underlying primitive is the same.

What was actually ingested

The full 26:09 transcript (556 ASR segments) at raw/videos/fix-karpathys-llm-wiki-with-knowledge-graph-claude-code-obsidian-infranodus.md. Auto-generated captions, fetched 2026-05-17 at --timeout 180000 after the initial 60-second attempt failed with the transcript-panel-did-not-render symptom (now a recognised long-format-video failure mode — see 2026-05-15 log entry on the MGI virtual-event retry pattern). 16-chapter structure preserved from the YouTube description. Light proper-noun verification at ingest time (Karpathy / InfraNodus / Obsidian / Claude Code spelled correctly in the body).

Cross-positioning with the wiki

The early-April LLM-Wiki cluster (one week post-Karpathy gist)

The wiki now holds three independent third-party engagements with Karpathy’s LLM Wiki pattern published within ~1-2 weeks of the upstream gist:

Source	Date	Vantage
OceanBase ex-brain	9 April 2026 (5 days post-gist)	Vendor implementation — ex-brain CLI tool with smart-compilation pipeline, embedded seekdb DB, MCP server integration
InfraNodus video (this)	11 April 2026 (7 days post-gist)	Architectural augmentation — KG-as-attention-direction layer, gap-analysis workflow, MCP server, Claude skill
Raju explainer	16 April 2026 (12 days post-gist)	Conceptual explainer + 10-dim RAG-vs-LLM-Wiki comparison

These three converge on a shared pattern: the LLM Wiki is the durable structural primitive; the question is what to layer on top (smart compilation / KG augmentation / RAG-bridge). The Nodus Labs piece is the most architecturally specific.

The KG-as-attention-direction claim

Prior wiki sources on knowledge graphs (SurrealDB, Manditereza HiveMQ, Leskovec) frame KGs as retrieval substrates (deterministic graph-traversal supplementing or replacing vector search). The Nodus Labs video adds a second use mode: KGs as attention-direction primitives — the gap-prompt is not retrieving anything; it’s telling the LLM where in its existing context to look harder. This is a structurally different role and worth a body note on the knowledge-graphs page.

Connection to harness engineering

The “living memory” framing — infranodus/ folder accumulating ontology graphs across sessions — is structurally parallel to Bratanic’s hooks-driven memory and Osmani’s AGENTS.md memory standard. The KG layer is to the LLM Wiki what the memory layer is to the harness: a compounding persistent-state substrate the model reads on every turn. This is the wiki’s first source that names KG-as-harness-memory-component explicitly, even though the speaker uses the word “wiki” rather than “harness.”

Connection to this very repository

This source recommends a pattern (KG augmentation of LLM Wiki + InfraNodus MCP server + per-session ontology graphs) that this repo does not currently use. The repo’s .graph.json (per CLAUDE.md §Graph) is a derived export, not an InfraNodus-style attention-direction primitive. Open architectural question for this repo: would adding an InfraNodus-style gap-analysis pass (probably as a lint operation per CLAUDE.md’s four operations, or as a new gap-analyse operation) materially improve the wiki’s value as a research substrate? The next harness-ratchet increment would be: make the wiki’s own structure-gaps queryable as a tool the agent calls during synthesis.

Named entities (this ingest)

• Nodus Labs — YouTube channel + company; creators of InfraNodus. First mention. Dangling. Channel-as-author convention applies.
• InfraNodus — knowledge-graph visualisation tool / product / web app + Obsidian plug-in + VS Code extension + MCP server. First mention. Dangling.
• Speaker (channel host) — self-identifies as the InfraNodus creator who has “been working on InfraNodus for the last 10 years.” The channel description / public-record attribution maps this to Dmitry Paranyushkin (InfraNodus founder); the wiki defers the named-person promotion to next ingest for direct verification.
• Andrej Karpathy — already in the wiki; cited as the upstream LLM Wiki gist author.
• Other passing entities (Obsidian / Claude Code / Cursor) already in the wiki ecosystem.

Source-quality notes

• Genre: practitioner tutorial / product-demo video by a tool vendor. Vendor-sponsored claim — per CLAUDE.md Lifecycle vendor-source rule, confidence is capped at 0.75.
• Empirical content: zero empirical study; full video is methodology demonstration on the speaker’s own research wiki. Generic claims are illustrated, not measured.
• Strength of the contribution: high despite the vendor genre. The KG-as-attention-direction framing is a genuinely new architectural primitive that the wiki’s other knowledge-graph sources do not name. The 16-chapter live-demo format is unusually transparent for a vendor video — viewers can replicate the demo end-to-end from the screen recording.
• Confidence: 0.72. Per heuristic: single source, +0.05 for live-demo transparency, vendor-sponsored cap at 0.75. Will rise to 0.80+ if an independent practitioner reproduces the gap-analysis workflow.