Definition
An LLM wiki agent is a knowledge management architecture where a language model — not the user — acts as the active maintenance agent for a structured personal knowledge base. Originated by Andrej Karpathy in a published gist. The core swap: instead of the human tagging, linking, organizing, and retrieving, the LLM does. The human’s role shifts to providing raw material and asking questions.
The minimal implementation requires three files:
- CLAUDE.md — schema file; defines categories, formatting rules, and standing operating instructions for the agent
- index.md — structured knowledge index; maintained and updated by the LLM at the close of every session
- log.md — append-only session diary; records what was discussed, what changed, what the system now knows
Why It Matters for the Newsletter
Technology & AI: This is the first practical implementation of Vannevar Bush’s 1945 Memex concept — an agent maintaining associative trails through a knowledge base on behalf of the user. The concept went from theoretical (1945) → technically possible but human-bottlenecked (1990–2022) → solved by LLMs (2023+).
Meta: This wiki operates on exactly this architecture. Every ingest, query, and lint session is an instance of the LLM wiki agent pattern in production.
Evidence & Examples
- Obsidian Was Never the Problem — explains the architecture and contrasts it with passive PKM systems (Obsidian, Notion, Building a Second Brain)
- This vault (Substack Research) — production implementation; CLAUDE.md defines schema, index.md catalogs pages, log.md records sessions
Tensions & Counterarguments
- Requires a capable LLM — the architecture is only as good as the model’s ability to maintain consistency across sessions
- Long-term drift: LLMs may introduce inconsistencies in entity names, link targets, or concept definitions without a human periodically auditing
- Context window limits: a very large wiki may exceed what the LLM can hold in one session; the index + log structure mitigates this but doesn’t eliminate it
- Dependency risk: the knowledge base is maintained by a proprietary model; if the model degrades or access is lost, the wiki still exists as static files but loses its active maintenance layer
Related Concepts
- PKM Failure Pattern — the problem this architecture solves
- Tech-State Conflict — LLM infrastructure is centralized and proprietary; relevant to who controls the tools