---
type: source-digest
topic: agent-wiki
status: active
created: 2026-04-06
updated: 2026-04-06
source_notes:
  - "[[clippings/26-04-06 - Github Repo - karpathyllm-wiki]]"
derived_from:
  - workspace-initialize
confidence: high
source_url: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f
---

# karpathy-llm-wiki

## Why This Matters

This is the conceptual seed for the whole workspace. It reframes knowledge work as compilation: raw sources stay immutable, the wiki becomes the maintained synthesis layer, and the agent behaves like the maintainer of that layer rather than a stateless answer engine.

## Key Takeaways

- Ordinary RAG and file-upload workflows often rediscover the same synthesis from scratch on every query.
- A maintained markdown wiki can act as a persistent, compounding knowledge layer between raw sources and answers.
- The durable artifact is not just the source archive; it is the network of synthesized pages, links, and contradictions.
- `index.md` and `log.md` are first-class operating files because they provide navigation and chronological memory.
- The schema file is what turns a generic LLM into a disciplined wiki maintainer.

## Entities

- Karpathy
- Obsidian
- `index.md`
- `log.md`
- schema / `AGENTS.md`

## Related Concepts

- [[research/agent-wiki/concepts/compiled-knowledge-layer|Compiled Knowledge Layer]]
- [[research/agent-wiki/concepts/obsidian-as-agent-ide|Obsidian as Agent IDE]]

## Open Questions

- What specific schema rules are required to keep a workspace trustworthy in this vault?
- When should a useful answer be saved as an output versus folded directly into concepts?
- At what corpus size does the file-and-index approach begin to strain?