本文由 AI 分析生成
建立時間： 2026-04-03 來源： https://x.com/jumperz/status/2039826228224430323
Summary
JUMPERZ comments on Karpathy’s LLM knowledge base workflow and extrapolates it to multi-agent systems. The core loop — dump research into a folder, let the model organise it into a wiki, ask questions, file the answers back in — compounds because every query enriches the base. Extended to agents, this means coordinators can maintain institutional knowledge across sessions instead of rebuilding context from scratch, and agents that own a local knowledge layer need good file organisation and index files rather than infinite context windows.
JUMPERZ 評論 Karpathy 的 LLM 知識庫工作流，並將其延伸至多代理系統。核心循環（倒入研究資料 → 模型整理成 wiki → 提問 → 將答案歸檔回去）具有複利效應，每次查詢都使知識庫更豐富。推廣至代理系統，意味著協調者可在跨 session 維護「機構知識」，而擁有本地知識層的代理不需要無限上下文窗口，只需要好的檔案組織和索引檔案。
Key Points
- The loop is the key insight: every query makes the wiki better; it compounds over time into a “second brain building itself”
- Applied to multi-agent systems: a coordinator that maintains institutional knowledge means each execution adds something back to the shared base — no cold-start per session
- Agents with their own knowledge layer don’t need infinite context windows; they need good file organisation and the ability to read their own index files
- This architecture is cheaper, more scalable, and more inspectable than stuffing everything into a single giant prompt
- The implication for agent design: knowledge persistence as a first-class concern, not an afterthought
Insights
The reframe from “personal tool” to “agent architecture” is the thread’s contribution. Karpathy described it as a productivity workflow; JUMPERZ generalises it: the same loop that makes a human researcher’s wiki compound also makes an AI agent’s institutional knowledge compound. The coordinator agent becomes a knowledge maintainer, not just a task router.
The “no infinite context needed” claim is important. Most discussions of scaling AI agents focus on extending context windows or improving retrieval. This suggests an alternative: invest in file organisation and indexing discipline, and a bounded context window is sufficient for arbitrarily large knowledge bases. The constraint becomes storage and search quality, not model capability.
“Way more inspectable” is a quiet but significant advantage: a markdown wiki is human-readable, version-controllable, and debuggable in ways that vector databases and latent model weights are not.
Connections
- LLM Knowledge Bases — the original Karpathy thread this post directly comments on
- agent-architecture
- institutional-memory
- knowledge-management
Raw Excerpt
agents that own their own knowledge layer do not need infinite context windows, they need good file organisation and the ability to read their own indexes. way cheaper, way more scalable, and way more inspectable than stuffing everything into one giant prompt.