RAG forgets; a wiki remembers

What it does This is essentially a workflow specification and prompt template packaged as an Agent Skill. It directs your coding agent to ingest sources into an immutable raw/ archive, compile the material into cross-linked markdown pages in wiki/, and answer future questions by citing those durable knowledge pages instead of re-processing originals on every query.

The interesting bit It treats the LLM as a curator, not just a retriever. The README draws a sharp contrast with RAG: synthesis happens during ingest and maintenance, so relationships and summaries improve over time rather than being rebuilt from chunks at query time.

Key highlights

• Three operations: Ingest (sources → raw/ → compiled wiki pages), Query (cited answers from existing wiki pages), and Lint (integrity checks, broken links, stale cross-references).
• Follows the open agentskills.io standard, installable in Claude Code, Cursor, Codex CLI, and OpenCode.
• Accepts web pages, papers, PDFs, blog posts, or pasted text; everything lands as markdown in raw/ before the LLM synthesizes it into durable wiki/ pages.
• Includes an append-only operation log and global index that the agent maintains automatically.
• Unofficial community implementation of Karpathy’s original gist, alongside related projects like lucasastorian/llmwiki.

Caveats

• The README cites a “production” knowledge base maintained daily since April 2026, a future date that appears to be a typo or placeholder.
• Codex CLI support requires manual copying to a skills directory rather than the single-command install that works for other tools.

Verdict Worth a look if you use Claude Code, Cursor, or Codex and want structured, long-term memory that improves as you feed it more sources. Skip it if you need broad retrieval across massive, static corpora where traditional RAG is a better fit.