LLM Wiki
An LLM Wiki is a persistent, structured knowledge base composed of Markdown files that is built and maintained by a large language model (LLM) agent rather than by humans directly. The concept was introduced by Andrej Karpathy in April 2026 through a 75-line GitHub Gist that went viral, receiving over 17 million views on X and 15,000 GitHub stars within a week. The core insight is that knowledge should be compiled once at ingestion time and kept current, rather than being re-derived from scratch on every query as in traditional Retrieval-Augmented Generation (RAG) systems. The architecture consists of three layers — an immutable raw sources layer, an LLM-maintained wiki layer, and a schema layer defining conventions — with three operations (ingest, query, lint) that enable knowledge to compound over time. Baidu Baike described the paradigm as shifting knowledge management from "repeated discovery" to "continuous growth."
By July 2026, the paradigm had been independently validated by four teams converging on the same three-layer architecture within months of Karpathy's post: Cognition's [[DeepWiki]] indexing 50,000+ code repositories, Factory's [[AutoWiki]] treating documentation as a build artifact, LangChain's OpenWiki expanding from code to personal knowledge, and Garry Tan's GBrain scaling to 146,646 pages with 66 autonomous cron jobs. LangChain's OpenWiki moved through a rapid July release cadence — launched on July 1, extended to general-purpose wiki memory with OpenWiki Brains on July 10, and integrated the Open Knowledge Format on July 16. Google Cloud released the Open Knowledge Format (OKF) v0.1 specification in June 2026, formally standardizing the LLM Wiki pattern as a vendor-neutral knowledge interchange format, followed by OKF v0.2 in July 2026 adding five trust signal fields for provenance, trust, freshness, lifecycle, and attestation in agent-generated knowledge. In spring 2026 the pattern was also productized as platform infrastructure: [[Anthropic]] shipped its Memory and Dreaming APIs for Claude Managed Agents, and [[OpenAI]] shipped Dreaming for ChatGPT, both implementing background consolidation of agent experience into persistent memory stores. The economic advantage of the pattern was empirically validated, with a controlled experiment demonstrating 84.6% token savings over RAG on sequential queries and calibrated 30-day projections of 53.7–81.3% cumulative savings. Practical validation on a 200,000-line Go codebase compiled 68 documentation files into 106 wiki pages and surfaced ten latent documentation contradictions in a single evening. Analysts have framed the pattern as the third stage of knowledge-base evolution, following low-end RAG and NotebookLM-class productized RAG, with the LLM Wiki adding persistent knowledge-structure sedimentation on top of automated retrieval. Karpathy's own vision extends the pattern to on-demand research: teams of LLMs could be dispatched to construct temporary wikis for each question, perform quality inspections, and produce comprehensive reports.
The ecosystem expanded into chat-native knowledge bases with Beever Atlas transforming team chat into structured Neo4j knowledge graphs, reasoning-based retrieval with OpenKB's PageIndex, agent memory integration through the Hermes LLM Wiki Memory Provider, and sustained production use demonstrated by Astro-Han's implementation with 94 wiki articles and 99 sources maintained daily since April. A new class of open-source tools now treats the llm wiki as the agent's long-term memory itself: nvk/llm-wiki provides an auditable research brain with 5–10 parallel research agents per command, ai-memex-cli packages Karpathy's pattern as a Git-backed skill-and-CLI toolbox, Memory Fort compiles observations from six agent runtimes into a curated wiki with BM25, semantic, and graph retrieval, and Open Second Brain embeds a deterministic nightly "dream" consolidation pass directly in an Obsidian vault. The pattern also reached end users through Obsidian community plugins that run entirely locally — LM Wiki and Auto LLM Wiki compile a user's existing vault into a queryable wiki without any cloud account — and was productized in mainstream office software when WPS Comate Wiki shipped an AI compilation engine that converts raw documents into a structured knowledge system. The nashsu/llm_wiki desktop application became the most popular open-source implementation, surpassing 15,000 GitHub stars with a Tauri v2 cross-platform design featuring two-step Chain-of-Thought ingest and knowledge graph community detection. The WiCER research algorithm (arXiv:2605.07068) addressed the compilation gap, demonstrating that blind LLM compilation fails catastrophically 53–60% of the time but recovers 80% of lost quality through targeted iterative refinement. Commercial products have begun combining the pattern with notebook-style research interfaces, most notably Knowly, launched in May 2026 as a closed-loop proactive AI pairing the LLM Wiki with NotebookLM-style learning flows.
Signed OKF added Ed25519 cryptographic signing for OKF bundles, and EQTY Lab with NVIDIA introduced hardware-rooted verifiable knowledge using confidential computing, extending trust from metadata to silicon-level proofs. Enterprise adoption accelerated with CLPS Incorporation's Project Athena, Jiran Software's Process Wisdom Star, and Qihoo 360's AI Knowledge Base 3.0. The first corporate LLM Wiki training course was organized in South Korea, and the Wikimedia community hosted its first Wiki AI Pre-Conference at Wikimania 2026 in Paris. Generative AI also entered Wikimedia editing workflows when the June 2026 EduWiki Workshop combined PETScan discovery tools with AI-assisted editing under the principle "Find with PETScan. Assist with AI. Verify as a human. Publish with confidence." The English Wikipedia community enacted a ban on unreviewed LLM outputs through a 40–2 vote, adopting the G15 speedy deletion criterion for immediate removal of AI-generated articles lacking human review. As enterprise architects cautioned, the pattern works best as a personal research tool or complementary layer rather than a full replacement for document management platforms. Hands-on critics reached a similar conclusion from the opposite direction, reporting that the basic pattern fails on large, multi-domain knowledge bases because index-based navigation collapses beyond a few hundred pages and contradiction handling is left unresolved. On the frontier of agentic software, Karpathy argued that as models make software output effectively free, a Jevons-paradox dynamic drives exponential growth in developer demand for hyper-specific tools and test suites, a trend that bears directly on the economics of automated knowledge compilation.