LLM Wiki
LLM Wiki is a design pattern in which an LLM incrementally transforms curated raw sources into a persistent collection of human-readable, interlinked pages and maintains that collection as sources and queries accumulate. In Karpathy's formulation, raw material remains immutable; the LLM writes summaries, entity and concept pages, comparisons, and indexes under a schema that defines ingest, query, and lint workflows. The name denotes a pattern, not one product, model, or storage engine.
Origin and context
Karpathy published a single-revision idea file on 4 April 2026 and explicitly left implementation details to users and their agents. The same name then appeared in independent research and software: one paper evaluates information loss during wiki compilation, another operationalizes the pattern as an agent-native retrieval system, and an unaffiliated repository implements its ingest, query, and lint loop. These are later instantiations, not co-originators of the term.
Why it matters
The pattern moves recurring synthesis and bookkeeping from every question into a maintained artifact. Pages can preserve explicit links, provenance, contradictions, and prior analyses, while useful answers can be filed back for later work. That can support cumulative research or team continuity. The benefit is conditional: independent studies show both promising structured traversal and a compilation gap in which a model can discard important facts while compressing sources.
Example
A research team could keep papers and meeting notes in an immutable raw directory. On ingest, an agent updates source, concept, and entity pages plus the index and log, preserving citations. A query reads several relevant pages, follows links, and files a useful synthesis back into the wiki. Periodic linting checks broken links, stale claims, and contradictions; human curators still choose sources and review consequential edits. Search can be added when the index is no longer sufficient.
How it differs
Retrieval-Augmented Generation
RAG is the broader pattern of supplying retrieved external evidence during generation. LLM Wiki precompiles and maintains a human-readable knowledge layer before a question arrives, but it may still search or retrieve from that layer. It is therefore neither a synonym for RAG nor proof that retrieval is unnecessary; Karpathy's contrast is with systems that repeatedly retrieve raw chunks without accumulating maintained synthesis.
GraphRAG
GraphRAG builds an entity graph and community summaries to support graph-aware retrieval and corpus-wide questions. An LLM Wiki can consist of ordinary Markdown pages and links governed by an editorial schema; it does not require a graph database, community detection, or GraphRAG's query pipeline. A system may combine both approaches without making them identical.
Maturity and evidence
Maturity is rated 3. The primary document specifies repeatable layers and operations, two unaffiliated papers analyze or instantiate the named pattern, and independent software implements its workflow. That is same-sense adoption beyond one post. It is not mature consensus: the evidence is only months old, the research is preprint evidence, implementations differ materially, and results from one concrete LLM-Wiki system do not validate the whole pattern.
Limits and open questions
A persistent artifact can preserve errors as effectively as knowledge. WiCER reports substantial information loss from blind compilation in its evaluation and improves it with diagnostic refinement. Wiki pages can omit fine detail, accumulate unsupported claims, become stale, or develop broken links and contradictions. Karpathy's moderate-scale observation is personal experience, not a general benchmark. Implementations should retain immutable sources and claim provenance, version changes, review consequential content, and evaluate compilation recall, update behavior, and answering quality against simpler RAG or full-context baselines.
Related terms
References
- LLM WikiAndrej Karpathy / GitHub Gist · 2026-04-04 · class A
- WiCER: Wiki-memory Compile, Evaluate, Refine Iterative Knowledge Compilation for LLM Wiki SystemsJuan M. Huerta / arXiv · 2026-05-08 · class B
- Retrieval as Reasoning: Self-Evolving Agent-Native Retrieval via LLM-WikiWeChat, Tencent / arXiv · 2026-05-25 · class B
- Beyond Memory: A Templated Substrate for Heterogeneous Collaborative Knowledge Work with LLM AgentsPriscila Saboia Moreira and Christopher R. Sweet / arXiv · 2026-05-29 · class B
- LLM Wiki: an independent local-first implementationYasas / GitHub · 2026-04 · class B
- Retrieval-Augmented Generation for Knowledge-Intensive NLP TasksFacebook AI Research, UCL, and NYU / NeurIPS · 2020-05-22 · class A
- From Local to Global: A Graph RAG Approach to Query-Focused SummarizationMicrosoft Research / arXiv · 2024-04-24 · class A
Last updated: 2026-09-07