ajeygore/llmwiki

LLMWiki is a lightweight personal knowledge base tool designed to give both humans and AI coding assistants a shared way to read and update the same information. A human can open the wiki in a browser and see a nicely formatted, readable page, while an AI agent works with the same content as plain Markdown files sitting in the codebase. It is a complete implementation of an architecture pattern originally described by Andrej Karpathy, whose gist is linked in the README. The core idea is to separate an AI agent's workspace knowledge, meaning its memories, rules, and custom skills, from the actual project source code. That knowledge instead lives in its own wiki repository, so the main project stays clean while a whole team can share the same collective context that their AI agents read from and write to. Setting it up means dropping the LLMWiki engine into an empty folder, either by cloning it and removing its git history so it becomes plain files in your own repository, or by adding it as a git submodule that stays linked to updates. A setup script then fills in the workspace with the pieces needed to run: agent instructions, an index page, and starter Markdown folders for content. Once running, you start a local server to browse the wiki as a dashboard in your browser, and there is also a command line search tool so both agents and humans can look things up directly from the terminal. To bring an AI coding assistant into the loop, you paste a welcome prompt from the dashboard into the assistant's chat, which tells it to read the agent instructions file and use built in workflows for adding, searching, and checking the wiki's content. If you already have existing project notes or custom skills, you can simply ask your coding agent to migrate them into the new wiki structure. Because the viewer works entirely in the browser using relative file requests, a finished wiki workspace can also be published as a read only site on GitHub Pages with no extra code, using a ready made publishing workflow the setup script creates automatically. The full README is longer than what was shown.