LLM Wiki: Agent-Maintained Knowledge Base
LLM Wiki: Agent-Maintained Knowledge Base
A personal knowledge management system where an AI agent continuously integrates new sources into a structured, interconnected wiki — inspired by Andrej Karpathy’s LLM Wiki concept. Rather than plain retrieval-augmented generation, knowledge compounds over time in browsable markdown files.
Project Overview
Traditional RAG re-derives structure on every query and discards insights. LLM Wiki instead maintains a growing vault of linked notes, where the agent ingests new material, updates summaries, tracks contradictions, and records an append-only activity log — making knowledge accumulate and become more useful over time.
Key Features
- Agent-Maintained Vault: LLM agent automatically integrates new sources into linked wiki notes
- Immutable Source Storage: Raw inputs are preserved separately from processed wiki content
- Summary & Contradiction Tracking: Agent updates summaries and flags conflicting information across notes
- Append-Only Activity Log: Full audit trail of all ingests and updates
- Obsidian Compatible: Vault format designed for graph visualization in Obsidian
- Lint & Maintenance Workflows: Built-in tooling to keep the knowledge base clean
Technical Implementation
- Storage: Markdown-based notes organized as a structured wiki
- Agent Framework: Designed to work with Claude or other LLMs as the maintenance agent
- Vault Format: Obsidian-compatible file structure with bidirectional links
- Organization: Separation between raw source documents and processed wiki content
Development Highlights
- Designed an append-only ingestion pipeline that prevents overwriting source material
- Implemented contradiction detection so the agent flags conflicting facts across notes
- Structured the vault to be browsable both programmatically and visually via Obsidian’s graph view
- Built around the insight that compounding knowledge outperforms stateless retrieval over time
Tech Stack
- Language: Markdown / plain text
- Agent: LLM-powered (Claude / Deepseek compatible)
- Visualization: Obsidian vault format
- Source Control: Git