On April 3, 2026, 📝Andrej Karpathy posted a detailed architecture description of "LLM Knowledge Bases" that landed as one of the most significant public signals in AI-native knowledge infrastructure.
What He Described
- Raw source documents (articles, papers, repos, datasets, images) are dropped into a
raw/directory - An LLM incrementally compiles them into a wiki — a flat directory of interlinked
.mdfiles - The LLM writes all wiki content: summaries, backlinks, concept articles, cross-references
- Schema lives in
AGENTS.md, which the LLM reads to understand how to organize and contribute - He uses Obsidian as a read-only IDE frontend, with Web Clipper for ingesting web articles
- Outputs from Q&A sessions are filed back into the wiki — explorations compound, the base grows from use
- Periodic health checks have the LLM scan for inconsistencies, impute missing data via web search, and surface candidates for new articles
At ~100 articles and 400,000 words, he found no need for RAG pipelines or vector databases — the LLM handles index files and document summaries natively at that scale. He closed with:
"I think there is room here for an incredible new product instead of a hacky collection of scripts."
Community & Market Signal
- Ole Lehmann (AI educator): packaging this for "normal people" is a massive product opportunity
- Lex Fridman confirmed he uses a similar setup for podcast research
- VentureBeat covered the "Karpathy Pattern" as a new architecture paradigm
- Commentary is already shifting from personal research wikis toward multi-agent orchestration — MythOS's actual territory
Karpathy's post describes a personal workflow shift that validates years of 📝MythOS development in a single public moment. He doesn't write throwaway posts about productivity workflows. He coined "vibe coding" in a single tweet and it reshaped the industry's vocabulary for a year. This post is more deliberate — a detailed architecture description published to millions of followers, naming both the pattern and the gap.
What he described is not a new idea. It is the MythOS thesis, articulated by one of the most trusted technical voices in AI:
- Token throughput shifting from code to knowledge — MythOS has been designed around this reallocation since V1
- Agents with a single source of truth — MythOS is the knowledge layer that keeps agents from "waking up blank"
- Flat, LLM-readable structure with a schema file — Karpathy's
AGENTS.mdis functionally identical to MythOS's memo-as-instruction model - Filed outputs that compound — MythOS memos accumulate as a living record of every session's output
- Health checks and self-healing — MythOS's architecture anticipates this as an agent-level maintenance task
The gap he named — "an incredible new product instead of a hacky collection of scripts" — is exactly what MythOS has been building toward since 2017. He did not know that. That's what makes it a flag, not a conversation.
This post arrived at the right moment. MythOS V3 is live and expanding. The communities layer is in development. The brand story site is in progress. Karpathy just handed us a credibility artifact from the highest-signal source in the ecosystem — the person who coined the vocabulary the entire field is using.
The framing opportunity: MythOS is the product Karpathy said needs to exist. That's not a stretch. It's accurate. And it's quotable.