Karpathy’s LLM‑Wiki demo
Andrej Karpathy showcased LLM‑Wiki powering Hermes Agents, which lets agents store local research and rapidly build skills with simple commands like '/llm-wiki <research x>'. The demo points to a lightweight pattern for agent-backed workflows that combine local memory with LLM tooling. (x.com)
Most people use a large language model like a search intern: you dump in files, ask a question, and it re-reads the pile every time. Andrej Karpathy’s demo showed a different pattern, where the model writes a living folder of markdown notes that stays on your machine and gets better after each round of research. (gist.github.com) Karpathy calls that folder an LLM Wiki, short for a large language model wiki. In his gist from early April 2026, he says the point is to build a “persistent, interlinked collection of markdown files” instead of rediscovering the same facts from raw documents on every query. (gist.github.com) The simple version is a cookbook. A normal retrieval system keeps all the groceries in bags and rummages through them at dinner time, while an LLM Wiki turns the groceries into labeled recipe cards you can reuse tomorrow. (gist.github.com) Karpathy’s own setup is plain: he keeps an agent open next to Obsidian, a markdown notes app, and lets the model edit pages while he browses the links and graph view. His line is that “Obsidian is the integrated development environment, the LLM is the programmer, the wiki is the codebase.” (gist.github.com) The news this week is that Nous Research wired that idea into Hermes Agent as a built-in skill. Hermes’ skills page describes Karpathy’s LLM Wiki as a tool to “build and maintain a persistent, interlinked markdown knowledge base,” then ingest sources, query the compiled notes, and lint them for consistency. (hermes-agent.nousresearch.com) Hermes Agent is an open-source assistant that already keeps local memory, searches past conversations, and can create or improve skills over time. Its documentation says conversations, memory, and skills are stored locally in `~/.hermes/`, and it can run against a local model endpoint instead of a cloud provider. (github.com) (hermes-agent.nousresearch.com) That makes the demo less like a flashy new model launch and more like a wiring diagram for everyday work. You can point an agent at papers, meeting notes, or a codebase, have it turn the mess into linked pages once, and then ask follow-up questions against the cleaned-up version instead of the whole mess again. (gist.github.com) (hermes-agent.nousresearch.com) The command style matters too. Hermes supports slash commands and quick commands, and the demo used a simple `/llm-wiki` flow to trigger research and update the local wiki, which is much closer to using a shell than using a software suite with ten menus. (hermes-agent.nousresearch.com) (x.com) The bet underneath all of this is that small, durable notes may beat giant, fragile context windows for a lot of real work. Karpathy’s gist argues that for many personal knowledge bases, the useful artifact is not the raw pile of files or the vector index behind it, but the compiled markdown layer the model keeps revising as new evidence arrives. (gist.github.com) If that pattern sticks, the winning agent may not be the one that remembers everything in one shot. It may be the one that leaves behind a neat folder of files another model, or another human, can pick up next week and keep building. (gist.github.com)
