pi-llm-wiki

Why we included this project

Most knowledge bases quietly rot: you save a PDF or a link, and it joins the pile you never open again. This project instead feeds raw sources (URLs, local documents, pasted text) to an LLM that maintains an interlinked wiki, reorganizing itself as new material arrives, following the pattern Andrej Karpathy outlined for a compounding knowledge base. Everything is written as plain Markdown with canonical frontmatter and stable citations, so the vault stays readable in Obsidian and portable through the Open Knowledge Format rather than locked into a proprietary export. The same knowledge model is also exposed over MCP, which means Claude Code, Cursor, and other agent clients can query the wiki directly instead of re-searching raw files on every question. For anyone building a long-lived second brain, that combination of durable, human-readable storage and agent access is what makes it worth trying.