This is Cang He's 576th original post. RAG and NotebookLM sound intelligent, but the problem is obvious: every time you ask a question, the model reassembles an answer from fragments, then scatters them again. Ask the same question a hundred times and it reassembles a hundred times. Knowledge itself never accumulates.
What this is
Based on Karpathy's publicly shared LLM Wiki architecture, the author uses Obsidian (a local Markdown note-taking tool) for storage and Codex or WorkBuddy as the execution Agent, building a three-layer personal knowledge base: the Raw layer stores clipped articles and chat logs as original material; the Wiki layer holds AI-organized concepts and topics; the Schema layer dictates how the AI archives, updates, and resolves conflicts.
The key shift is that AI moves from one-off Q&A to long-term maintainer. When new material arrives, the Agent first scans existing pages, supplements what's already there, creates new concept pages, and when it encounters disagreement, preserves the source, timestamp, and scope of applicability. The knowledge base retains its history of changes rather than resetting to zero every time.
Industry view
Supporters see this as a paradigm shift in personal knowledge management. Obsidian's local Markdown format is naturally suited for Agent work: it can read and write files directly, use Git to track changes, and the backlink and graph features help users surface core nodes and orphaned pages. DeepSeek V4 Flash offers strong cost-performance but lacks multimodal support; Kimi K3 and Doubao-Seed-Evolving are the domestic multimodal alternatives.
The objections deserve equal attention. The original author himself strongly recommends against using GPT 5.6 Sol inside Codex, citing brutal token consumption. This means the system demands significant model context capacity and reliable tool-calling stability; real-world runtime and cost are not necessarily friendly. A more fundamental question: when the rules in AGENTS.md are handed to AI for self-execution, how do you audit errors accumulated over long-term maintenance? If the Agent quietly writes biased interpretations into the Wiki layer, can a human catch it in time? The burden of knowledge management shifts from "manually organizing" to "reviewing what AI organized," which for many people is not actually lighter.
Impact on regular people
For enterprise IT: The natural extension of this approach is department-level knowledge bases — distilling meeting minutes, product docs, and customer communications into evolvable structures instead of letting RAG re-retrieve and re-stitch every time.
For individual professionals: Technical roles willing to wrestle with Obsidian and Git can try it first; for non-technical users who just want to manage reading notes, NotebookLM remains the lowest-friction entry point.
For the consumer market: Obsidian-style tools will continue to ride the "data sovereignty" sentiment wave, but people who can fully run an Agent workflow end to end remain a minority. The real barrier is designing the Schema rules, not installing the tools.