I built a custom kernel — the LLM never touches the disk
Letting an LLM directly read/write flat .md files is a nightmare for scale. If you let an LLM rewrite markdown to maintain a complex graph, it inevitably hallucinates wiki links, breaks formatting, or drops paragraphs when context gets tight. Markdown isn't a database. I built a custom kernel instead. In my stack, the LLM never touches the disk — it only extracts structured semantic signals. The Python backend compiles those into typed nodes/edges and commits them to an SQLite FTS5 graph DB.
— @lauratom, 2026-04-04, via Discord.