LLM Wiki
Andrej Karpathy's pattern for AI-maintained personal knowledge bases: a three-layer architecture where LLMs incrementally build and maintain persistent markdown wikis from raw sources. The human curates sources and asks questions; the LLM does the grunt work of summarizing, cross-referencing, filing, and bookkeeping. The key distinction from RAG: knowledge is compiled once into structured pages and kept current, rather than re-synthesized from raw documents on every query.
Key Quotes
"You never (or rarely) write the wiki yourself -- the LLM writes and maintains all of it. You're in charge of sourcing, exploration, and asking the right questions."
Key Themes
This is the pattern this wiki itself implements. Three layers: raw sources (immutable), the wiki (LLM-owned markdown pages), and a schema document (CLAUDE.md) defining structure and workflows. Three operations: ingest (new source triggers multi-page updates), query (LLM searches wiki, synthesizes answers), lint (periodic health checks).
The insight that makes it work: LLMs don't tire or forget, so the tedious maintenance that kills human-maintained wikis (updating cross-references, maintaining consistency, catching contradictions) becomes tractable. The human does the interesting work (choosing what to read, asking questions); the LLM does the boring work (filing, linking, formatting).
The community response -- dozens of implementations addressing contradiction detection, token compression, hierarchical routing, provenance tracking -- validates the pattern's generativity. See Elements of Agentic Systems Design for the formal framework (particularly Memory and Artifacts elements).
Critical Analysis
Strong: the pattern works. This wiki exists because of it. The three-layer separation is clean and the operations (ingest/query/lint) are intuitive. Using Obsidian's graph view alongside the LLM agent is a practical workflow.
Weak: cascading updates from single changes can be expensive (tokens and time). Referential integrity without rigorous infrastructure is fragile -- "dead links" are a real risk as the wiki grows. The pattern assumes a single user with a single LLM maintaining everything; multi-user or multi-agent scenarios need transactional guarantees the markdown-file approach doesn't provide.
The deepest limitation: the LLM's synthesis is only as good as the sources. Garbage in, well-formatted garbage out. The human's curation role is understated in the original proposal.
A sharper dissent comes from Keep AI Out of Your (Obsidian) Vault: for a personal second brain, the problem isn't only garbage-in — it's that AI-written notes and AI-made connections aren't yours at all. mehdio argues the pattern automates away the very act of linking an idea yourself, which is where insight comes from, and that "your notes will be prompts or libraries for AI tomorrow, but not if you generate them." The pattern is safest as retrieval and cross-referencing; generation is where it starts to eat the human's thinking.
Sources: llm-wiki Last updated: 2026-05-14