Garry Tan’s GBrain sharpens agent knowledge bases

Strong release because it treats knowledge as a maintained system, not just a pile of notes.

• –The top/bottom page split is the real insight: it separates belief state from provenance, which is much better for agent use than pure append-only logs.
• –The scaling advice is useful because it matches what many markdown vaults hit in practice: grep works until it very much does not.
• –Hybrid retrieval with pgvector makes sense once you have many entity-centric pages and fuzzy recall matters more than exact filename recall.
• –The overnight consolidation loop is the most operationally interesting part; it turns the knowledge base into a self-healing system instead of a passive archive.
• –If you are already using topic MOCs plus grep, the likely inflection point is less about raw file count and more about how many entity pages need semantic retrieval.
• –The main risk is complexity creep: once you add search, vectors, and autonomous maintenance, the quality of the belief/timeline split depends on disciplined updates.