In early 2026, Andrej Karpathy - the AI researcher known for his work at Tesla and OpenAI - posted a sketch of a personal knowledge system on X. The concept: dump bookmarks and raw content into a folder, let a language model compile it into a linked wiki with citations, then search and browse it in Obsidian. The post got 17 million views.
A developer has now shipped the whole pipeline as a single NPM package called CacheZero.
The tool has four components: a Chrome extension that saves tweets, articles, and YouTube videos; a Hono web server with LanceDB for semantic search (searching by meaning, not just keywords - so "productivity software" finds results about "task management tools" even without the exact phrase); Claude Code to compile saved content into linked wiki pages with citations; and Obsidian for browsing the resulting vault.
Install is one command: npx cachezero.
This is a developer's weekend project made public - expect rough edges and incomplete documentation. But it's a working end-to-end implementation of the exact system Karpathy described, including the Obsidian integration and citation linking. The GitHub traction it's picking up suggests the original idea resonated for a reason: most people who bookmark things constantly lose them to the abyss of unread tabs and uncategorized folders.
Anyone frustrated by the gap between saving things from the web and actually finding them later will find this worth 20 minutes to test.