How Andrej Karpathy Builds His Second Brain: The “LLM Wiki” Architecture (Claude Code + Obsidian)

Most personal knowledge management systems become digital graveyards. We bookmark articles, clip web pages, and save research PDFs with the noble intention of revisiting them, only to amass thousands of fragmented files that are never opened again. In early 2026, former Tesla AI Director and OpenAI founding member Andrej Karpathy introduced an alternative paradigm: the “LLM Wiki”. Instead of expecting humans to manually organize notes or relying on superficial RAG vector queries, an autonomous coding agent acts as a persistent personal librarian, compiling raw information into a durable, compounding knowledge base.

The 15,000-Document Graveyard: Why Traditional Notes and Basic RAG Fail

Modern knowledge workers suffer from information obesity. According to cognitive science research, human memory deteriorates along the Ebbinghaus forgetting curve within 48 hours unless active recall occurs. To combat this, builders hoard bookmarks and markdown notes, but manual tagging and linking quickly collapse under scale.

When generative AI emerged, the initial industry answer was standard RAG (Retrieval-Augmented Generation): chuck thousands of raw PDFs and notes into a vector database, compute cosine similarity over chunks, and inject top matches into the system prompt. In practice, this approach fails for complex knowledge synthesis because:

• Raw Noise: Unstructured web clippings contain boilerplate, outdated specs, and promotional fluff that dilute the context window.
• Fragmented Synthesis: Answering a high-level question requires synthesizing five related concepts across six months of work, not retrieving three isolated text chunks.
• Zero Compounding: Every query starts from scratch. The system never learns or refines its mental model between conversations.

The Andrej Karpathy Paradigm: The LLM as an Active Librarian

During his Sequoia Capital address and subsequent technical memos, Andrej Karpathy formulated a foundational thesis for agentic engineering:

“The LLM writes and maintains the wiki; the human reads and asks questions. The wiki is a persistent, compounding artifact.”

Instead of treating the LLM as a transient chatbot, you grant it write access to a structured local filesystem. When you discover new research, code architectures, or business strategies, you drop the raw artifact into an ingestion queue. The AI librarian digests the material, identifies connections to existing knowledge, writes new comprehensive encyclopedia articles, and updates cross-references across your entire vault.

The 3-Tier Architecture: Raw, Wiki, and Schema

The system operates across three strictly delineated layers inside a standard project root:

my-second-brain/
├── raw/                      # Immutable Source Material (Read-Only for LLM)
│   ├── ai-engineering/       # Web clips, whitepapers, interview transcripts
│   ├── business-strategy/    # Competitor teardowns, revenue models
│   └── architecture/         # System designs, database schemas
├── wiki/                     # Compiled Knowledge Base (Full LLM Ownership)
│   ├── index.md              # Global Master Index (Category + Summary + Date)
│   ├── log.md                # Append-Only Audit Trail of all modifications
│   ├── ai-engineering/       # Distilled, interconnected Markdown articles
│   │   ├── agentic-workflows.md
│   │   └── memory-systems.md
│   └── business-strategy/
└── SKILL.md                  # Schema Layer & Operating Protocol

1. raw/ — The Immutable Evidentiary Layer

This directory stores source documents exactly as captured. The AI agent has read-only access and is programmatically barred from altering or deleting files here. This guarantees that your original reference material remains pristine and auditable.

2. wiki/ — The Compounding Knowledge Layer

This is where the magic lives. The AI agent possesses complete ownership over this directory. Articles are organized into single-level topic subdirectories (e.g., wiki/ai-engineering/agentic-workflows.md). Every article adheres to clean Markdown formatting, complete with bidirectional links ([[article-name]]) that native viewers like Obsidian render into visual neural graphs.

3. SKILL.md — The Behavioral Protocol

The schema file defines the agent’s exact operational rules, formatting templates, and quality constraints. Whether executing via Claude Code CLI, Cursor Agent, or custom autonomous containers, the schema governs how inputs are evaluated and compiled.

The Core Lifecycle: Ingest, Query, and Automated Linting

The LLM Wiki operates through three primary workflows:

1. Ingest (The Self-Healing Backlink Phenomenon)

When new material is added to raw/, the agent executes an ingestion pass. Crucially, the agent does not merely append a new document. It scans the existing wiki/ directory to determine how this new information alters prior conclusions.

If you ingest a benchmark on new inference speeds, the agent updates the historical LLM economics article, amends comparison tables, links the new research, and logs the change in wiki/log.md. Your past notes are continuously updated and corrected by your AI librarian without manual intervention.

2. The Grounding Invariant (Zero-Hallucination Fact Verification)

To eliminate synthetic hallucinations, the Karpathy architecture enforces a strict invariant: every load-bearing fact, metric, date, and verbatim quote in wiki/ must exist verbatim in the linked raw/ source files. During compilation, the agent verifies exact string matches. A lightweight Python verification script can re-check the entire wiki in seconds by cross-referencing literals against source files.

3. Query & Periodic Linting

When you query your second brain, the agent answers by citing verified wiki articles rather than guessing or searching the open web. Meanwhile, automated linting sweeps detect broken internal links, orphan articles with zero backlinks, conflicting dates, or missing index entries.

Step-by-Step Implementation with Claude Code & Obsidian

1. Create your Vault: Download Obsidian and initialize a new local vault named my-second-brain.
2. Install Claude Code: Open your terminal and install the official Anthropic Claude Code CLI runtime (npm install -g @anthropic-ai/claude-code or native terminal installer).
3. Deploy the Skill: Add the MIT-licensed community specification (e.g., Astro-Han/karpathy-llm-wiki) or write your custom CLAUDE.md inside the root directory.
4. Clip and Ingest: Use the official Obsidian Web Clipper extension to save high-value articles into raw/. Run claude "ingest raw/new-article.md" and watch your AI librarian compile, cross-link, and catalog the knowledge in real time.

The Paradigm Shift for Solo Builders

Building leverage in the generative AI era is not about memorizing chat prompts; it is about building durable information infrastructure. By separating raw evidence from synthesized knowledge and assigning an autonomous agent to maintain your records, you convert transient reading time into a permanent, self-updating capital asset.