LLM Wiki - Karpathy's open-source, self-running personal knowledge base

What is LLM Wiki?

LLM Wiki, a self-running personal knowledge base launched by Karpathy, breaks through the limitations of traditional RAGs that "retrieve from scratch with every query." LLM Wiki guides LLMs to proactively maintain a structured Markdown Wiki through schema files, "compiling" raw data into a persistent knowledge body with cross-references and contradiction annotations. Supported by a three-tier architecture (raw data/Wiki/Schema), knowledge continuously accumulates and grows with use, avoiding repetitive derivations. Humans focus on asking questions and curating, while LLMs handle maintenance, making Obsidian the IDE, the Wiki the codebase, and the LLM the programmer.

Main functions of LLM Wiki

• Knowledge ingestionPlace the original materials (PDF, articles, etc.) into raw/ The table of contents, LLM automatically extracts key information, writes summaries, and updates entity pages, concept pages, and related cross-references in the Wiki (one piece of information may trigger 10-15 page updates).
• Intelligent QueryAnswer questions based on compiled Wiki content, and output supports multiple formats such as Markdown, comparison tables, Marp slides, and matplotlib charts.
• Health Inspection (Lint)It regularly and automatically checks the knowledge base for contradictions, outdated conclusions, isolated pages, and missing citations, and proposes new research directions.
• Index maintenanceAutomatic maintenance index.md(Table of Contents, a vector search alternative to traditional RAGs) and log.md(Timeline Log) can support medium-sized knowledge management (approximately 100 articles/400,000 words) without the need for a complex database.

How to use the LLM Wiki

• Environmental preparationInstall Obsidian as the reader and prepare Claude Code and other LLM agents as maintainers.
• Initialize directory: Create containing raw/(Original data)wiki/(Generated content) and CLAUDE.md(Schema configuration) folder structure.
• Configure SchemaCopy the contents of Karpathy Gist to the Agent, which will then generate configuration files that define the Wiki structure, page format, and workflow.
• Input dataPlace the new file in raw/ The directory is then processed by the Agent, and the LLM will automatically extract key information and update the summary, entity pages, and cross-references in the Wiki accordingly.
• Query ArchivesAsk the Agent a question to get a comprehensive answer based on the Wiki, and instruct the best answers to be saved back to the Wiki as a new page to achieve continuous knowledge accumulation.
• Health InspectionRegularly instruct the Agent to check the Wiki for contradictions, outdated conclusions, and isolated pages to maintain a clear knowledge base structure and consistent content.
• Browse and useOpen in Obsidian wiki/ View the latest updates in folders in real time, click on bidirectional links and use the graph view to explore knowledge connections.

Key information and usage requirements of LLM Wiki

• natureThis allows LLM to proactively maintain a continuously accumulating structured knowledge base, rather than the traditional RAG's "retrieving from zero for every query".
• ArchitectureThree-layer structure: raw/ (raw data, read-only) → wiki/ (Markdown generated by LLM) → Schema (such as CLAUDE.md rule configuration).
• Core Operations: Ingest (automatically updates multiple pages by inputting data) → Query (answers based on Wiki and can be archived) → Lint (contradictions and deficiencies in regular health checks).
• Key mechanismsKnowledge is compiled once and kept fresh; cross-references and contradictory annotations are persisted and become richer with use.

LLM Wiki's core advantages

• Knowledge compoundingBreaking away from the limitations of traditional RAGs, which rely on "retrieving from scratch and discarding after use" for every query, Wikis are persistent and cumulative products—new information is automatically integrated into the existing knowledge network, and cross-references and contradictory annotations are continuously accumulated.
• Zero maintenance burdenLLM handles all the tedious maintenance work, automatically updating 10-15 related pages, marking discrepancies between old and new data, and maintaining indexes and cross-references. Humans can focus solely on curating and asking questions, eliminating the need for manual organization.
• Medium-sized projects exempt from infrastructure constructionWithin a scale of 100 articles/400,000 words, efficient retrieval can be achieved using only Markdown index files, without the need for vector databases, embedded models, or complex RAG architectures, significantly reducing the technical threshold.
• Dynamic self-reinforcementHigh-quality questions and answers can be archived back to the Wiki as a new page with one click. The exploration process itself continuously enriches the knowledge base, forming a flywheel effect of "growth through use".

LLM Wiki project address

• GitHub repository: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f

Comparison of similar products on LLM Wiki

Comparison Dimensions	LLM Wiki	NotebookLM (RAG representative)	Mem (AI Notes Representative)
Knowledge Model	Active compilation, Continuous accumulation	Passive retrieval, Each time, piecing it together from scratch	AI automatically organizes. But there is no explicit compilation layer.
Main maintenance	LLM full-time maintenance (Automatic updates for multiple pages) (Note: Contradictions)	No maintenance Each temporary search	AI black box processing, Rules are difficult for users to intervene in.
Core Architecture	Three-layer separation: raw/ (read-only) wiki/(written by LLM) Schema (rules layer)	Single layer: Original data directly as Search source	Double layer: Notes content + AI-related
Indexing mechanism	Markdown index file (Medium-sized vector library)	Vector Embedding + Similarity Search	Private Algorithm (Black box)
Customizability	White box control: pass CLAUDE.md Custom page format, Update rules, workflow	Fixed process: Unable to intervene in the retrieval and generation logic	Black box restrictions: Only has on/off function. Unable to define maintenance rules
Output sedimentation	High-quality answers are acceptable Archive back to Wiki Create a new page	The answer disappeared Dialogue with History	Can be saved but Lack of structure Archiving mechanism
Technical threshold	Markdown only Files + Obsidian	Relying on Google Infrastructure	Relying on third parties cloud services

Application Scenarios of LLM Wiki

• Personal growthTrack your goals, health, mental state, and self-improvement journey; organize diaries, articles, and podcast notes to build a structured cognitive map of yourself.
• academic research: To conduct in-depth research on a specific topic for several weeks or months, reading papers and reports, and gradually building a comprehensive knowledge system that includes evolutionary arguments.
• Reading CompanionWhen reading books, organize them chapter by chapter, create related pages for characters, themes, and plot lines, and ultimately form a complete knowledge network similar to Tolkien Gateway.
• Enterprise/team knowledge baseIt integrates Slack conversations, meeting minutes, project documents, and customer communication records, and LLM automatically maintains and updates its internal Wiki in real time, requiring no additional manual effort.