Overview
What this is
The compact read before the technical details.
A living knowledge vault managed by Claude that transforms raw sources into interconnected wikis of concepts, questions, and themes. Inspired by Karpathy's wiki pattern, it ingests documents, articles, and notes, then uses Claude's reasoning to extract key ideas, identify relationships, and surface unresolved questions. The result is a self-organizing knowledge base that evolves as you add more material.
Capabilities
What it actually does
The useful parts, pulled out of the paragraph wall.
Agent-operated knowledge vault following Andrej Karpathy's LLM wiki pattern — raw sources stay separate from synthesized concepts, creating a compounding searchable thinking surface
Strict 2-source rule prevents concept page proliferation: a theme must appear in at least two independent sources before it graduates from candidate to concept page
Clean separation of concerns: inbox → immutable source notes → synthesized wiki → archived residuals, with agents owning all downstream maintenance while humans gate what enters the system
Full lifecycle management including vault graph generation (Mermaid), health audits, keyword drift detection, and an append-only operation log for provenance
Implementation
Technology with jobs attached
Names are less useful than responsibilities. This is what each piece is doing.
Claude API / Codex
Primary LLM agent that runs ingest, compile, ask, lint, and graph commands
Python
Fetch and processing scripts for YouTube transcripts, ArXiv papers, and web articles
trafilatura
Web article content extraction with boilerplate removal for URL ingestion
youtube-transcript-api
Fetches YouTube video transcripts for video source ingestion
Obsidian
Recommended vault interface for browsing and editing the Markdown knowledge base
Mermaid
Generates vault connection graphs and architecture diagrams inline in Markdown