Karpathy's LLM Wiki post put agent knowledge into the public conversation. The principles are solid: knowledge separate from the source corpus, incremental building, extensibility across domains. The space needs the attention.

Most of the implementations that followed treat the post as a spec: per-file ingestion into markdown articles, prompt-based maintenance rules, periodic lint passes. Some went further and boiled it down to a Claude Code skill. The energy is great, but these approaches shift the maintenance headache into a new form rather than solving it.

Here's where they break and what works instead.