LLM Wiki Pattern

A pattern for building personal knowledge bases where an LLM agent incrementally builds and maintains a persistent, interlinked wiki between the user and their raw sources. The user curates and questions; the LLM does all writing, summarizing, cross-referencing, and bookkeeping.

This is the pattern this Second Brain follows. The schema (CLAUDE.md) is the local instantiation.

Three layers

1. Raw sources — immutable input documents
2. The wiki — LLM-generated, interlinked markdown
3. The schema — config (e.g. CLAUDE.md) defining structure, conventions, and workflows

Three operations

• Ingest — read source, extract, file as a source page, update entity/concept pages, append to log
• Query — answer from the wiki; file non-trivial answers back as query pages so they compound
• Lint — periodic health check: contradictions, stale claims, orphans, missing cross-refs

Why it differs from RAG

RAG retrieves and re-stitches fragments on every query. The wiki compiles knowledge once and keeps it current. Cross-references and contradictions are pre-resolved. Synthesis is persistent.

Why the LLM has to do the maintenance

The hard part of any knowledge base is bookkeeping at scale: 15 cross-references per source, contradictions across dozens of pages, summaries that need refreshing. Humans drop the habit; LLMs don't. This is the part Memex couldn't solve in 1945.

Tooling that supports the pattern

• Obsidian for viewing (graph, wikilinks, plugins)
• Claude Code (or another agent) for writing
• Optional: Marp, Dataview, qmd

Cognitive-science backing: the graveyard-of-ideas trap

Independent corroboration from a learning-theory source — How To Learn Anything So Fast (theMITmonk):

"So many of us spend more time designing and organizing our digital system — pages, folders, views, tags, databases — and so little time actually connecting ideas in our head. But a graveyard of ideas is still a graveyard."

The MITmonk's diagnosis is exactly the failure mode this pattern is designed to dodge. His proposed fix is human discipline (the TRAP Framework); the LLM Wiki Pattern's bet is structural — the connection work the human won't sustain is exactly what an LLM agent will. The two are complementary rather than competing: the LLM maintains the connections (graph), the human does the Test/Perform steps (retrieval and judgment).

Caveat: the wiki can still feed the graveyard pattern if the user's only loop is "read it." See Fluency Illusion — the vault is recognition-optimized; retrieval pressure must come from outside (drafting, deciding, explaining cold).

Executive-friendly framing

When introducing the pattern to non-technical peers, the three operations rename cleanly (see Sunil's Second Brain Email to IT LT (2026-06-06)):

Schema vocabulary	Executive vocabulary
Ingest	Capture anywhere — two seconds of effort per source
File / cross-link	Let the agent do the librarian work — twice a day
Query	Ask the vault, not a blank search bar — it cites your own notes

The renaming matters because the schema words ("ingest", "wikilinks", "frontmatter") signal a developer's pattern; the executive words signal a learning system — which is what makes peers actually try it.