LLM Wiki Pattern#

A pattern for building personal knowledge bases where the LLM incrementally builds and maintains a persistent, interlinked wiki from raw sources. Proposed by andrej-karpathy. This wiki is a running instance of this pattern.

Core Insight#

Wiki > RAG. RAG rediscovers knowledge from scratch on every query — no accumulation. An LLM-maintained wiki compiles knowledge once and keeps it current. Cross-references, contradictions, and synthesis compound with every source added.

Three-Layer Architecture#

Raw Sources (immutable)  →  Wiki (LLM-maintained)  →  Schema (CLAUDE.md)
   you add these              LLM writes these         you + LLM co-evolve

Raw sources: Immutable documents. LLM reads, never modifies.
The wiki: LLM-generated markdown. Summaries, entities, concepts, comparisons, synthesis. LLM owns entirely.
The schema: Instructions telling the LLM how the wiki works — structure, conventions, workflows. The key configuration file.

Three Operations#

Ingest: Process a source → create/update 10-15 wiki pages → update index → append to log
Query: Read index → find relevant pages → synthesize answer → optionally file back as new page
Lint: Health-check for contradictions, stale claims, orphan pages, missing cross-references, data gaps

index.md: Content-oriented catalog. LLM reads first on every query. Works at moderate scale (~100 sources, ~hundreds of pages) without embedding infrastructure.
log.md: Chronological append-only record. Parseable with grep "^## \[" log.md.

Why It Works#

Humans abandon wikis because maintenance burden grows faster than value. LLMs don’t get bored, don’t forget cross-references, can touch 15 files in one pass. The maintenance cost is near zero.

The human’s job: curate sources, direct analysis, ask good questions, think about meaning. The LLM’s job: everything else.

Intellectual Lineage#

Related to Vannevar Bush’s Memex (1945) — a personal, curated knowledge store with associative trails between documents. Bush’s vision was closer to this than to what the web became: private, actively curated, connections as valuable as documents. The part he couldn’t solve was who does the maintenance.

Connections to the Ecosystem#

The schema (CLAUDE.md) is effectively a skill in the agent-skills-standard sense — instructions that shape agent behavior for a domain
Index-first navigation is a form of context-management — selective loading, same principle as progressive disclosure
Could scale to multi-agent-orchestration: separate agents for ingest, query, lint, with the wiki as shared state
claude-code’s CLAUDE.md and auto memory are the native implementation substrate

Contrast: NotebookLM#

notebooklm shares the “human curates sources, AI processes them” principle but takes a different path:

Session-oriented vs. cumulative: NotebookLM notes live within a session/notebook; the LLM Wiki compounds knowledge permanently across all sources.
Interactive exploration vs. structured ingest: NotebookLM encourages open-ended conversation then pinning; the wiki follows a defined ingest → entity → concept → cross-reference workflow.
Manual cross-reference vs. automatic wikilinks: NotebookLM’s “Suggest Related Ideas” is on-demand; the wiki maintains links as a structural invariant.

Both validate the core insight: grounding AI in curated sources produces better knowledge work than open-ended generation.

Formalization: Open Knowledge Format (OKF)#

In June 2026, google-cloud-platform published the open-knowledge-format — an open specification that formalizes the LLM Wiki pattern into an interoperable standard. OKF takes what Karpathy described as a personal practice and provides the missing interoperability layer:

Minimal contract: Only type is required. Reserved filenames (index.md, log.md) provide navigation conventions.
Producer/consumer independence: Who writes the wiki and who consumes it are decoupled. One LLM’s output can be queried by another LLM, browsed by a human, or ingested by a catalog.
Format, not platform: No SDK, no cloud account, no runtime required. Just files.

OKF explicitly cites Karpathy’s gist and the broader pattern convergence (Obsidian vaults, AGENTS.md, metadata-as-code repos) as prior art. The key observation: all these instances look alike but aren’t designed to cooperate. OKF provides the agreed-upon conventions that make them interoperable.

This wiki vs. OKF: Our schema already follows the OKF shape (markdown + YAML frontmatter + cross-links + index.md + log.md). The main delta is wikilinks (**page**) vs. standard markdown links ([page](/path/to/page.md)). We use a richer set of conventions (source/entity/concept/analysis types, strict ingest workflow, log format) that are compatible with but go beyond OKF’s minimal spec.

Llm Wiki Pattern