I have harnessed llmwiki based on Obsidian to Damoang.
Introduction : https://damoang.net/ai/5251
Usage example : https://damoang.net/ai/5616
Reintroduction (What is LLMwiki)
Unlike RAG, which searches for knowledge for each question, llmwiki organizes knowledge once when the source is input and continuously updates it. It mimics the structure of human memory by classifying information into four levels: working memory → episodic memory → semantic memory → procedural memory. Each page is assigned a trust level, last confirmation date, and decay rating.
What's New
The wiki, which was previously used only within Obsidian Vault, has been made publicly available on npm as an MCP server. You can now use it outside of Obsidian Vault. You can access the wiki while working with Codex CLI, Cursor, or modifying code in other repositories.
claude mcp add --scope user llmwiki -- npx -y obsidian-llmwiki-mcp --root
The server automatically creates 8 code snippets for Claude Code, Codex, Gemini CLI, agy, Cursor, Windsurf, Claude Desktop, and VS Code. It is read-only and does not write anything to the vault.
Some Numbers (Same A/B Test Conditions)
Factual briefing queries: 50.6k → 22.7k tokens (−55%), tool calls 14 times → 1 time
Procedural/how-to queries: No improvement (approximately 39.5k → 38k). Procedures inherently require viewing multiple pages, so scoping did not reduce them.
2nd MCP server call: 164ms → 26ms (process cache) for a 1,000-page synthesized vault
Every commit checks if the Python original script and output are byte-identical (56 fixtures + 20 real vaults + 420 differential fuzzing, 3 OS CI)
Not Yet Implemented or Verified
The agy client's tool selection is unstable even when rule snippets are added.
Gemini CLI and Cursor have only been verified for configuration file merging, not actual use.
NFD filenames on macOS may not align with NFC queries.
Related Links
▶ Original source: https://skillmaru.hell0world.net/space/global/llmwiki-harness
▶ Original source: https://github.com/cookyman74/llmwiki-harness
