llm-wiki-kit

An MCP server that implements Karpathy's LLM Wiki pattern - persistent, LLM-maintained knowledge bases that compound over time.

Instead of RAG (rediscovering knowledge from scratch on every query), the LLM incrementally builds and maintains a structured wiki with interlinked markdown files, cross-references, summaries, and synthesis that get richer with every source you add.

The tedious part of maintaining a knowledge base isn't the reading or thinking, it's the bookkeeping. Updating cross-references, keeping summaries current, noting contradictions, maintaining consistency. LLMs are perfect for this. You curate and direct. The LLM does everything else.

Imagine you are researching a new and complex technology like LLM speculative decoding. Instead of reading 10 papers and taking manual notes, you use llm-wiki-kit to let your agent build a state map over time.

1. Human: drops 3 PDFs into raw/
2. Human: "Analyze these papers and update the KB. Pay special attention to KV cache optimizations."
3. Agent (via MCP):

• Calls wiki_ingest for each paper
• Calls wiki_write_page to create concepts/speculative_decoding.md
• Calls wiki_write_page to update synthesis/cache_strategies.md and link it to the papers
• Calls wiki_lint to ensure the new "Draft Model" concept is cross-referenced with existing "Inference" pages

Two weeks later, you start a fresh chat session in Cursor or Claude Code. You do not need to re-upload the papers or re-explain what you learned. You ask:

> "Based on our research so far, which draft model architecture is most efficient for Llama 3?"

Your agent calls wiki_search, reads the synthesis pages it wrote earlier, and answers from accumulated evidence:

> "Based on the compiled evidence in your KB, the Eagle architecture is currently leading b