Freedium
Short summary
Karpathy proposes replacing on-the-fly retrieval with an LLM-maintained markdown wiki that ingests sources, compiles structured pages, and self-updates through ingest, query, and lint cycles. This approach builds a persistent, compounding knowledge base without vectors or re-retrieval, though it currently lacks enterprise controls.
- Karpathy's "LLM Wiki" replaces vector-based RAG with an LLM that reads sources once and maintains a persistent, self-updating markdown wiki instead of re-retrieving chunked fragments per query.
- The system runs three operations—ingest (add new material), query (answer from the wiki, then save new answers back into it), and lint (scheduled scans for outdated info, broken links, and gaps)—so every interaction compounds the knowledge base rather than starting fresh.
- It directly targets two RAG failure modes: stateless retrieval that repeats embedding work every query, and chunking that destroys document context.
- The idea echoes Vannevar Bush's 1945 Memex concept of an associative, hyperlinked personal knowledge library, though it currently lacks enterprise-grade controls.
Full summary
Andrej Karpathy sketched out a new way to build personal knowledge bases he calls “LLM Wiki.” Instead of chopping documents into vectors every time you ask a question, this pattern has the LLM read raw sources once, extract key points into a structured markdown wiki, and keep that wiki up to date. No external vector database. You drop PDFs, articles or transcripts into a raw/ folder. The agent reads them, writes summary pages, updates concept and entity pages, and maintains a master index. Contradictions get flagged, cross-references get created automatically.
The system runs three operations. Ingest pulls in new material, updates all relevant wiki pages and the index. Query searches the wiki’s existing pages to answer your questions and then turns useful answers into fresh wiki entries—so every interaction enriches the knowledge base. Lint is a scheduled health check: the LLM scans for outdated info, broken links, orphaned pages and fills gaps. Over time, the wiki grows richer and more consistent without re-retrieving fragmented context.
Karpathy’s pitch tackles two chronic RAG problems: stateless retrieval that repeats the same embedding dance on every query, and chunking that destroys context by slicing up documents. LLM Wiki treats knowledge as a codebase—immutable sources feed into a live “program” of markdown code that accumulates understanding. The result is coherent articles, entity pages and indexes you can query instead of scattered text fragments.
He even ties it back to Vannevar Bush’s 1945 Memex vision—an associative, hyperlinked library of one’s own records. Markdown files and LLM agents aren’t new, but this framing flips RAG on its head: build once, update continuously, and let your queries compound learning rather than starting from scratch each time.