Your AI re-derives everything it knows from scratch on every question. One architecture decision stops it.

Source: Nate’s Newsletter Date: 2026-04-22 URL: https://natesnewsletter.substack.com/p/your-ai-re-derives-everything-it

Summary

Critiques Karpathy’s viral AI wiki (100K+ bookmarks) by identifying the architectural decision most readers missed: write-time vs query-time knowledge synthesis. Karpathy’s wiki pre-compiles knowledge when information enters the system. Open Brain processes and connects at query time. The choice determines cost profiles, failure modes under load, trust placement, and scalability. Key warning: “a neglected wiki is more dangerous than a neglected database” — write-time synthesis creates hidden decay risks. Proposes hybrid architecture combining structured storage (Open Brain) with compiled synthesis (Karpathy).

Implications

Extends the context portability thread into knowledge architecture. The write-time vs query-time fork maps directly onto agent memory design: Gemini CLI’s four-tier memory (write-time extraction) vs Claude Code’s session-scoped context (query-time synthesis). The hybrid recommendation validates what the agent vendors are converging on — extract durable knowledge at write-time, but keep query-time synthesis for connecting it. The “neglected wiki” warning is the maintenance cost of persistent memory that no vendor has addressed yet.