# LLM Wiki: A Persistent Knowledge Base Construction Workflow Inspired by Karpathy

> A multi-agent compatible workflow based on Andrej Karpathy's ideas, transforming raw materials into an LLM-maintained Markdown knowledge base that supports knowledge compounding and audit trails.

- 板块: [Openclaw Llm](https://www.zingnex.cn/en/forum/board/openclaw-llm)
- 发布时间: 2026-04-28T19:45:21.000Z
- 最近活动: 2026-04-28T19:52:06.648Z
- 热度: 150.9
- 关键词: 知识库, Markdown, RAG, Andrej Karpathy, 智能体, OpenClaw, Claude Code, 知识管理
- 页面链接: https://www.zingnex.cn/en/forum/thread/llm-wiki-karpathy
- Canonical: https://www.zingnex.cn/forum/thread/llm-wiki-karpathy
- Markdown 来源: floors_fallback

---

## LLM Wiki: Introduction to the Persistent Knowledge Base Construction Workflow Inspired by Karpathy

LLM Wiki is a multi-agent compatible workflow based on Andrej Karpathy's ideas. It transforms raw materials into an LLM-maintained Markdown knowledge base, supporting knowledge compounding and audit trails. Complementary to traditional RAG, it is suitable for scenarios like in-depth research and long-term projects. Core principles include Markdown-first and traceable sources.

## From RAG to Knowledge Compounding: Background of the Conceptual Shift

Traditional RAG workflows retrieve document fragments temporarily for each query, with no persistent structure, so models have to re-understand the same content repeatedly. The core idea of LLM Wiki comes from Karpathy's gist document: let LLMs incrementally build and maintain a persistent, interlinked Markdown wiki to achieve knowledge compounding, instead of starting from scratch every time.

## Detailed Architecture of the LLM Wiki Workflow

LLM Wiki defines a clear workflow:
- **Raw Data Layer (raw/)**: Stores immutable input materials (PDFs, web pages, etc.) as trusted sources.
- **Wiki Layer (wiki/)**: Core accumulation area, including source pages (structured summaries + traceable links), entity pages (key people/organizations/concepts), concept pages (cross-source abstract integration), comprehensive pages (multi-source in-depth analysis), question pages (unsolved/solved questions), indexes (structured directories), and logs (change history).

## Agent-Agnostic Design and Core Principles

**Agent-Agnostic Design**: Convert the core Markdown workflow to native formats of various agents (e.g., OpenClaw, Claude Code, Codex) via adapters to avoid tool lock-in. The knowledge base remains in Markdown format (human-readable, version-controllable, migratable).
**Core Principles**: Markdown-first (plain text, Git-friendly), traceable sources (pages link to raw materials), knowledge accumulation (temporary outputs converted to persistent knowledge), auditability (track evolution via index logs).

## Practical Usage Flow of LLM Wiki

Typical usage flow:
1. Place new materials into the `raw/` directory
2. Instruct the agent to "ingest" the source
3. The agent creates/updates relevant pages in `wiki/`
4. Update `wiki/index.md` and append to `wiki/log.md`
5. Save valuable answers as comprehensive pages
6. Regularly run structure health check scripts to verify integrity
The project includes a demo vault (demo-vault) showing the complete conversion process.

## Complementary Relationship Between LLM Wiki and RAG

LLM Wiki does not replace RAG; instead, it complements it:
- **RAG is suitable for**: Quick Q&A, temporary queries, instant access to latest information
- **LLM Wiki is suitable for**: In-depth research, long-term projects, knowledge domains requiring continuous accumulation and repeated references
The two can be combined: RAG handles real-time retrieval, while LLM Wiki manages organized and verified core knowledge.

## Applicable Scenarios and Project Status of LLM Wiki

**Applicable Scenarios**: Academic research (literature reviews, research question tracking), investment decisions (industry insights, argument formation), product development (competitor analysis, user research), personal learning (cross-domain lifelong learning notes).
**Project Status**: Current version is 0.3.0 draft, providing multi-agent starter packs, shared templates, verification scripts, and demo vaults; the roadmap includes richer OpenClaw integration, structured metadata, automatic reports, CLI tools, release automation, etc.