LLM Wiki Agent — source summary
An MIT-licensed, open-source coding-agent skill (by SamurAIGPT) that implements the
llm-wiki pattern directly: drop sources into raw/, tell the agent to ingest, and
it maintains a persistent interlinked wiki. Works with Claude Code, Codex, OpenCode,
and Gemini CLI; no API key or Python setup needed. Delivered via Telegram, ingested
2026-05-29. README in raw/llm-wiki-agent.md.
What it is
The most literal implementation of llm-wiki seen in the wiki — and a near-twin of
this very wiki. Structure: raw/ (sources) → wiki/ with index.md, log.md,
overview.md (living synthesis), and sources/, entities/, concepts/,
syntheses/ subfolders; plus a graph/ (graph.json + interactive graph.html).
Operations: ingest / query / lint / build graph, exposed as natural-language
triggers and Claude Code slash commands (/wiki-ingest, /wiki-query, /wiki-lint,
/wiki-graph) — making it an instance of agent-skills. Ingests many formats
(PDF/DOCX/PPTX/…) via markitdown auto-conversion.
Notable specifics
- Auto-created entity & concept pages updated on every relevant ingest; contradiction flags raised at ingest time, not at query; query answers can be filed back as synthesis pages (the “explorations compound” idea from llm-wiki-gist).
- Two-pass knowledge-graph build: a deterministic pass parses all wikilinks
→
EXTRACTEDedges; a semantic pass infers implicit relationships →INFERRED(with confidence) orAMBIGUOUS. Community detection clusters topics. This is a middle point between GBrain’s zero-LLM typed edges and pure manual linking.
Where it sits in the family
A lightweight counterpart to gbrain: both self-maintaining LLM knowledge bases, but llm-wiki-agent is markdown-only, runs inside an existing coding agent, needs no DB or API key — whereas GBrain adds Postgres, vector search, a daemon, and scale. On the llm-wiki automation spectrum it sits near the LLM-Wiki end, slightly automated.
Related
llm-wiki · llm-wiki-gist · gbrain · knowledge-graph · agent-skills