AgnosticBrain adds a protected layer to stop AI from overwriting your manual edits

A knowledge management project called agnosticBrain has introduced a three-layer ownership system to address a blind spot in Andrej Karpathy's LLM Wiki pattern, which Karpathy published in April 2026. The original pattern splits a personal knowledge vault into raw sources (read-only for the LLM) and a wiki directory (fully maintained by the LLM), but assumes the human never edits the wiki directly. In practice, users do refine AI-generated notes by hand, and subsequent LLM ingests can silently overwrite those edits. AgnosticBrain solves this by adding a third directory called curated/, where manually refined notes are stored and treated as strictly off-limits for any AI agent or automated operation. Users move a note into curated/ via a /curate command, permanently protecting its contents from being altered during ingests or any other pipeline task.