LLM Wiki Pattern

The LLM Wiki pattern is an approach to personal knowledge management where an LLM incrementally builds and maintains a persistent wiki from raw sources. Instead of re-deriving answers from documents at query time (as RAG systems do), the LLM reads each new source, extracts key information, and integrates it into an existing structured knowledge base — updating entity pages, revising summaries, flagging contradictions, and adding cross-references.

The wiki is the persistent artifact. It compounds over time: every source ingested and every question asked makes it richer. The human curates sources and directs exploration; the LLM handles summarizing, cross-referencing, and bookkeeping.

This second brain wiki follows the LLM Wiki pattern. See llm-wiki in meta/ for the full description.

The pattern contrasts with project-centric systems like PARA, which organize by actionability rather than building interconnected knowledge. It also differs from progressive-summarization, which relies on the human to distill notes over repeated visits — here, the LLM does the distillation in a single ingest pass.

Structurally, the wiki is a file-based graph database — markdown files as nodes, wikilinks as edges, folders as taxonomy. Its purpose, as Rumproarious argues in filesystem-is-graph-database, is best understood through the lens of context-engineering: the wiki exists so future LLM conversations can be seeded with the relevant prior state.

Huashu's obsidian-ai-complete-guide documents a production implementation of this pattern: 2,000+ files, 55 custom Skills, a CLAUDE.md routing system, and seven practical workflows. He frames the wiki as a "compiler" rather than retriever — instead of re-deriving answers each time, the LLM compiles raw material into persistent, cross-linked articles.

By mid-2026 the pattern had escaped personal knowledge management and become a storage decision that shipping products make. codealmanac applies it to a codebase (markdown under almanac/, reviewed in Git, fed from agent transcripts on a schedule); agentmemory extends it with confidence scoring, lifecycle and hybrid search as an agent-memory server; openhuman mirrors its memory tree into a Karpathy-style Obsidian vault. llm-wiki-as-agent-memory covers the convergence and what these implementations are actually buying — auditability and portability rather than better recall.

The pattern produces wiki-shaped content; what to do with it is a separate decision. One natural publication target is a digital-garden — the format and the LLM Wiki share the same Markdown-plus-wikilinks substrate, so a static-site generator like quartz can render the wiki as-is.