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 as a pattern for personal knowledge management that shifts the role of the LLM from a passive retrieval-and-generation tool to an active knowledge compiler and curator. Karpathy described the division of labor by analogy: "Obsidian is the IDE, LLM is the programmer, wiki is the codebase"[^c1]. 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[^c2]. Karpathy identified the historical bottleneck that had prevented Vannevar Bush's 1945 Memex vision for eighty years: "the hardest part of maintaining a knowledge base has never been reading—it's been bookkeeping"[^c10].
The architecture consists of three layers: an immutable raw sources layer containing original documents and materials, a wiki layer of LLM-generated and maintained Markdown pages with cross-references, and a schema layer (such as a CLAUDE.md file) that defines the conventions and workflows for the LLM to follow. Three core operations—ingest, query, and lint—govern the system's lifecycle, enabling knowledge to compound over time as each new source and query enriches the existing knowledge base.
The LLM Wiki pattern has generated widespread community adoption, with dozens of open-source implementations, commercial products, and academic papers emerging within months of its introduction. LangChain released OpenWiki, an open-source agent and CLI for generating and maintaining codebase documentation that implements the pattern[^c3]. A preregistered academic study on progressive disclosure found that LLM Wiki retrieval achieves non-inferior answer quality while reducing costs by 33–55% across all tested regimes[^c4]. Tencent's LLM-Wiki agent-native retrieval system outperformed seven baselines including HippoRAG 2, LightRAG, and GraphRAG by 2.0–8.1 F1 points on multi-hop QA benchmarks[^c5]. Google Cloud formalized the underlying structure as the Open Knowledge Format (OKF) in June 2026, establishing a vendor-neutral specification for representing knowledge as Markdown files with YAML frontmatter[^c8]. Industry analysts at LangChain identified "wiki memory" as an emerging standard pattern for agent long-term memory in June 2026[^c11].
The pattern has also been recognized as transforming documentation from a human convenience into load-bearing runtime infrastructure for coding agents, where stale documentation can directly cause incorrect code changes[^c6]. Academic research has simultaneously quantified fundamental limitations: the LLMpedia study found that GPT-5-mini achieves a verifiable true rate of only 74.7% on Wikipedia-covered subjects when generating articles from parametric memory, and the WiCER study documented a 53–60% catastrophic failure rate when LLMs blindly compile raw documents into wiki pages without iterative refinement[^c12]. Enterprise and community implementations have converged on a compiled truth plus append-only timeline page structure, where the executive summary is rewritten as knowledge evolves while the evidence trail remains immutable[^c7]. Scaling guidelines have been formalized through community experience, addressing the primary failure mode of the wiki becoming a context bottleneck as it grows[^c9].