LLM Wiki

LLM Wiki is a knowledge management system based on an AI-native architecture, inspired by Andrej Karpathy's "LLM Wiki" concept. It positions the AI Agent as an active "Knowledge Gardener" that maintains a structured, human-readable Markdown knowledge base with bidirectional linking by deeply digesting raw materials.

Core Philosophy

Unlike traditional RAG (Retrieval-Augmented Generation), LLM Wiki emphasizes stateful knowledge evolution:

• AI Gardener Mode: AI is responsible for information extraction, summarization, entity recognition, and cross-document referencing, rather than simple fragment retrieval.
• Structured Storage: Knowledge is digested and settled into clearly categorized Markdown files, facilitating both human reading and continuous AI editing.
• Transparent & Visible: All knowledge storage consists of standard Markdown files, which can be managed directly as a local knowledge base.

Usage

You can drive the entire knowledge base lifecycle through dialogue with the AI Agent:

1. Ingest

Place your raw materials (such as PDFs, web clippings, or rough notes) into the raw/ directory.

• Example Instruction: "Process the new files in raw/ and extract the core concepts from them."
• Action: AI will read the materials, generate summaries in wiki/sources/, and create or update entries in wiki/concepts/ or wiki/entities/.

2. Query

Ask deep questions based on the existing structured knowledge.

• Example Instruction: "Based on the records in the wiki, help me summarize the relationship between [Concept A] and [Concept B]."
• Action: AI will prioritize retrieving indices and related files under the wiki/ directory to provide a systematic answer with citations.

3. Lint

Periodically perform a "health check" on the knowledge base.

• Example Instruction: "Check if there are any orphan pages or contradictory definitions in the wiki."
• Action: AI will scan the integrity of bidirectional links and propose optimization suggestions.

Directory Structure

• raw/: Raw Material Layer. Stores immutable source files (PDFs, articles, notes, etc.). AI has read-only access.
• wiki/: Knowledge Layer. The knowledge encyclopedia generated and maintained by AI.
  • wiki/concepts/: Definitions and in-depth analysis of concepts.
  • wiki/entities/: Profiles of people, organizations, or tools.
  • wiki/sources/: Summaries and metadata of the raw materials in raw/.
  • wiki/index.md: The main directory and index of the knowledge base.
  • wiki/log.md: Historical audit log of AI operations.
• outputs/: Temporary outputs or exported reports.