LLM Knowledge Bases (LLM-curated personal wikis)

Definition

A pattern for maintaining a continuously accumulating, structured personal or team knowledge base using an LLM. Unlike RAG, which synthesizes from scratch on every query, the LLM performs the organizing, cross-referencing, and maintenance itself, so that knowledge is made permanent in the pages.

Representative implementation: Karpathy LLM-wiki gist — a tree of markdown files plus ingest/query/lint workflows.

Why It Matters

• The meta-principle of this system. CLAUDE.md is a concrete application of this pattern.
• An attempt to solve the fundamental limitation of RAG ("rediscovery on every query, no accumulation")
• Vannevar Bush's 1945 Memex vision — the LLM solves the "who maintains it" problem

State of the Art (2026)

• Karpathy's LLM-wiki gist (2026-04) — 3-layer (Raw sources / Wiki / Schema), markdown + frontmatter, ingest/query/lint cycle
• Karpathy himself: "a large fraction of my recent token throughput is going less into manipulating code, and more into manipulating [knowledge]"
• Tooling ecosystem: Obsidian (graph view), Marp (slides), qmd (local BM25 + vector + rerank search)

Architecture (Karpathy pattern)

Layer 1 — Raw Sources (immutable)

• Snapshots of original articles/PDFs/data
• The LLM only reads them

Layer 2 — The Wiki (living markdown)

• Entity, concept, model, person, paper, and trend pages
• Cross-references (wikilinks) managed automatically

Layer 3 — The Schema (operating rules)

• A schema document such as CLAUDE.md defines the wiki structure and authoring rules
• The LLM consults it on every task

Workflows:

• Ingest — add a new source → update 10-15 related pages simultaneously
• Query — read the index → read related pages → answer (good answers are saved as new pages)
• Lint — check for contradictions/orphans/dead links

Why This Project Uses It

This ai-trend-notifier is itself an application of the LLM-wiki pattern. Adapted to the domain of AI trend tracking:

• the wiki is structured into 6 categories: entities/models/concepts/people/papers/trends
• ingest daily / lint weekly / synthesize quarterly
• daily outputs (briefs/daily/) are separate but cross-linked with the wiki

Sub-concepts / Related

• Vannevar Bush Memex page TBD
• Obsidian / PKM (Personal Knowledge Management) general theory TBD

Open Problems

• Automatic contradiction resolution — how far can the LLM decide autonomously? At what point is a user decision required?
• Schema drift — as topics diversify, the schema can break
• Information half-life — criteria for judging old pages as stale
• Token cost — the rising cost of reading the entire wiki vs. the accuracy trade-off of index-driven selective reads

Key Sources

• Karpathy gist — the original source
• people/karpathy — the originator
• Memex (Vannevar Bush, 1945) — Wikipedia

Related Concepts

• Agents (LLM Agents) (indirect — as a knowledge management agent)
• RAG (alternative pattern), PKM (umbrella concept) TBD

Notable Quote

"the LLM is rediscovering knowledge from scratch on every question. There's no accumulation." "the wiki is a persistent, compounding artifact"

— Karpathy, LLM-wiki gist

Personal Significance

The very existence of this page is the system's self-reference. By tracking its own operating principles in the wiki, the system achieves meta-consistency.