Your LLM's reading list, now with compound interest

What it does

pi-llm-wiki is an extension for the pi coding agent and oh-my-pi that ingests URLs, PDFs, and pasted text into a structured Markdown wiki. It follows Andrej Karpathy’s LLM Wiki pattern to summarize sources, interlink concepts, and maintain canonical pages that accrete meaning as you add material. The output is an Obsidian-compatible vault that compounds knowledge instead of treating every query as a fresh one-shot retrieval.

The interesting bit

The project treats knowledge as infrastructure with actual guardrails: it blocks direct edits to raw sources and generated metadata, runs deterministic linting for orphans and broken links, and layers a personal ~/.llm-wiki/ vault beneath any project wiki so your notes follow you across directories. It also natively writes Open Knowledge Format (OKF) v0.2 documents, betting that your second brain should outlive any single tool.

Key highlights

• Four-layer pipeline: immutable raw packets → source pages → canonical wiki pages → deterministic indexes and activity logs.
• Layered recall: searches personal and project vaults simultaneously, deduplicating and prioritizing personal results.
• Mechanical maintenance: built-in linting for coverage gaps, contradictions, and stale captures, plus an optional auto-update watch.
• OKF v0.2 native: portable documents with canonical frontmatter and stable source citations, readable by legacy tools.
• MCP server: exposes the same wiki to Claude Code, Cursor, and Windsurf via stdio transport.

Caveats

• Agent working-memory tools (trajectory capture and skill distillation) are disabled by default and require explicit opt-in.
• The shared “company” vault layer is listed as future work.
• The full slash-command experience is designed around pi and oh-my-pi; the MCP server offers access, but not the complete workflow.

Verdict

Best suited for developers already using pi or oh-my-pi who want their research to accumulate into a browsable, interlinked archive. If you are not in that ecosystem, the MCP server provides a narrower on-ramp, and the Obsidian compatibility may still be useful as a standalone knowledge formatter.