The article examines how the ease of building AI systems shifts the challenge from construction to defining what is built, using ontology’s five perspectives and the Viable System Model to diagnose Karpathy’s LLM Wiki, revealing strengths in entity‑level design and gaps in process and purpose.
The article analyzes why AI agents forget, then compares three memory architectures—RAG, LLM Wiki, and GBrain—detailing their strengths, weaknesses, scalability, latency, compounding knowledge, and autonomy, and offers guidance on choosing the right approach for different use cases.
LLM Wiki, proposed by Andrej Karpathy, replaces repetitive RAG retrieval for personal agents with a three‑layer markdown‑based knowledge base that separates raw sources, curated wiki pages, and schema constraints, enabling durable, auditable memory, structured updates, health checks, and a hybrid Wiki‑RAG workflow.
This article walks through using the youdaonote‑llm‑wiki skill to automatically ingest a set of Markdown articles into a cloud‑synced Youdao Note knowledge base, generate structured Wiki pages, perform cross‑document queries with citations, and keep the repository up‑to‑date, while comparing it to Karpathy's original script‑based approach.
Unlike typical RAG that discards knowledge after each query, the open‑source tools llm_wiki and SkillClaw let you continuously compile a personal knowledge base and evolve AI agents by incrementally storing documents and session‑derived skills, complete with multi‑step processing, community‑tested benchmarks, and cross‑platform support.
The article details a five‑layer architecture for a sustainable local knowledge vault—Inbox, Raw Sources, Extracted Text, Distilled, and Archive—explaining each layer’s purpose, the import workflow from tools like Evernote, and the supporting Python scripts that automate ingestion, extraction, and governance.
LLM Wiki is an open‑source, cross‑platform desktop application that transforms documents into an organized, interlinked knowledge base; unlike traditional RAG it incrementally builds a persistent wiki, offers a three‑layer architecture, Obsidian compatibility, and provides step‑by‑step installation and quick‑start guidance.
This guide details two ways to integrate Claude with Obsidian—direct MCP connection (about 30 minutes to set up) and a Karpathy‑inspired LLM Wiki workflow (raw→wiki→drafts)—including full configuration steps, a CLAUDE.md template, three core commands, and tips for turning raw notes into structured, publishable drafts.
