From the person who popularised the term "context engineering", the brain half of the problem at personal scale. Not RAG, which re-derives answers from raw documents on every query, but an LLM-maintained wiki: a persistent, compounding artifact where synthesis happens at ingest, cross-references are already there and contradictions have already been flagged. Three layers (immutable sources, maintained wiki, a schema that turns the model into a disciplined maintainer rather than a chatbot) and three operations (ingest, query, lint). Permissions are barely mentioned, which is exactly the pattern this list keeps finding: scale it to an organisation and the schema layer is where the access model would have to live.