Summary

This framework enables AI coding agents to build and maintain an Obsidian wiki, applying Andrej Karpathy's "LLM Wiki" pattern. It compiles and continuously updates interconnected markdown knowledge bases, making them accessible and manageable by various AI agents, thereby avoiding repetitive LLM queries and RAG.

How It Works

A central .skills/ directory houses agent-executable markdown skills. setup.sh configures agent-specific bootstrap files and symlinks these skills. AI agents ingest diverse sources (text, PDF, images), extract concepts, resolve conflicts, cross-link information, and track provenance (extracted, inferred, ambiguous) within an Obsidian vault for knowledge integrity.

Quick Start & Requirements

Clone the repository, cd into it, and run bash setup.sh. The script prompts for your Obsidian vault path and globally installs skills. Requires an AI coding agent capable of file interaction (e.g., Claude Code, Cursor, Windsurf, Codex, Gemini, Copilot). Vision-capable models are needed for image ingestion. Optional: QMD for semantic search.

Highlighted Details

• Universal Agent Integration: Unified skill management across multiple AI agents via setup.sh.
• Delta Tracking: Efficient updates using .manifest.json to process only changed content.
• Provenance Tracking: Distinguishes extracted, ^[inferred], and ^[ambiguous] claims for transparency.
• Multimodal Ingest: Processes text, PDFs, and images (requires vision models).
• Knowledge Graph Analysis: Generates insights (hubs, bridges) and exports graph data (JSON, GraphML, Neo4j, HTML).
• Tiered Retrieval: wiki-query optimizes performance by checking titles/summaries before full page reads.
• Optional QMD Integration: Enables concept-level semantic search and source discovery.

Maintenance & Community

The project is described as "early." No specific details on maintainers, community channels, or roadmap are provided in the README.

Licensing & Compatibility

License information is not specified in the README, potentially impacting commercial use or integration.

Limitations & Caveats

The framework is "early," requiring further development for skill intelligence, deduplication, and handling large vaults. Multimodal ingest needs specific vision-capable models. The absence of a specified license is a significant adoption consideration.