Obsidian

A markdown-based knowledge management tool that stores notes as plain .md files in a local folder, with a visual graph view of the links between them. The reference viewer in Andrej Karpathy's original LLM wiki pattern — used for browsing the LLM-maintained markdown files and exploring page relationships visually.

Key facts

• Type: Markdown note editor + graph view
• Storage: Plain markdown files in a local folder (no database)
• Distinguishing feature: Visual graph view of inter-page links — the inspiration for this wiki's own graph view at /wiki/wiki-graph/
• Role in Karpathy's pattern: The "IDE" for browsing the LLM wiki. The LLM edits files; the human browses the graph ^[src-002]
• Memory-system role: Roberts frames Obsidian as the long-term memory option when the user wants readable, editable markdown files, backlinks, visual graphs, and no vector infrastructure ^[src-059].
• Voice-agent role: Simmons uses a local Obsidian MCP/API plugin so a GPT Realtime 2 voice assistant can create, search, read, and write notes in a vault ^[src-104].

What it does

Obsidian treats a folder of markdown files as a "vault" and renders them as an interconnected notebook. Any [[wikilink]] inside a file creates a bidirectional edge between two pages, which is then visible in the graph view — a force-directed visualisation of the whole vault.

In LLM Knowledge Bases (Karpathy pattern), Obsidian plays the role of the browsing layer while Claude Code (or another LLM) acts as the author. The LLM edits the files; the human uses Obsidian to explore the result. This division of labour is the core of Karpathy's LLM wiki pattern: the LLM is the programmer, Obsidian is the IDE, the wiki is the codebase ^[src-002].

In practice, Obsidian is optional. The wiki is just markdown in a folder — any editor (VS Code, Cursor, Claude Code's terminal UI) works just as well for reading the files. Obsidian adds value specifically because of its graph view, which makes relationships between notes browsable in a way a directory listing cannot ^[src-002].

In Roberts's three-layer memory system, Obsidian is contrasted with Pinecone. Obsidian is best when the user wants to open, read, edit, and graph memory by hand; Pinecone is better when the goal is semantic search across many records, transcripts, or long documents ^[src-059].

Roberts's Hermes Agent OS demo uses Obsidian as a shared readable vault for Hermes Agent. The point is not only note-taking: the vault becomes part of the cross-surface memory bridge, so a mobile assistant can answer from the same durable context that the desktop agent environment can inspect ^[src-079].

Simmons's GPT Realtime 2 demo adds a voice layer to the same idea. Once Obsidian exposes a local MCP/API endpoint, the voice agent can create a note, open documentation in the browser, and paste or save content into the vault by spoken command ^[src-104].

How this connects to Robin's work

Obsidian is useful here because it illustrates a broader production-AI principle: durable AI systems need readable memory, explicit structure, and inspectable links between sources, concepts, and decisions.

Related starting points:

• LLM Knowledge Bases (Karpathy pattern) — the markdown wiki pattern that makes Obsidian valuable as a browsing layer.
• AI Knowledge Graph — the public graph view of this site's AI wiki.
• Enterprise AI in Production — the operating-model layer behind production AI systems.

Related

• See also: Claude Code, Hermes Agent, LLM Knowledge Bases (Karpathy pattern), AI Memory Operating System, Cross-Harness Memory Bridge, Conversation Wrap-Up Memory, Expert Knowledge Index, Pinecone, Model Context Protocol (MCP), Voice-Driven Desktop Agents

Source references

• ^[src-002] Robin Cartier — "Karpathy's LLM Knowledge Base: A Practitioner's Verdict" (2026-04-08)
• ^[src-013] Nate Herk — "Build & Sell Claude Code Operating Systems (2+ Hour Course)" (2026-05-01)

– Demo: 36 YouTube transcript nodes visualised in Obsidian with backlinks; Obsidian Web Clipper Chrome extension used to ingest the AI2027 article directly into the raw/ folder, triggering Claude Code to create 23 wiki pages.

• ^[src-059] Jack Roberts — "This Memory System just 10x'd Claude Code" (2026-05-03)
• ^[src-079] Jack Roberts — "Hermes Agent just got 10X Better (Agentic OS)" (2026-05-15)
• ^[src-104] Pat Simmons – "GPT Realtime 2 Can Now Run Your Entire Computer (Just Your Voice)" (2026-06-17)