Maester

AI agents forget. Worse, they misremember — and a GTM system that drafts outreach from misremembered facts is a liability, not leverage. My agents needed one memory they could all read and write, where a fact's provenance survives every rewrite and an operator can always answer: where did this claim come from?

Maester is a three-layer knowledge architecture served through Garry Tan's gbrain as the engine. The layers are the contract: raw is append-only ground truth — API responses, scrapes, transcripts — never edited, never "cleaned." Structured is deterministic normalization: no opinions, no hallucinated fields, null for unknown. Insights is where judgment lives — opinionated, but every claim traces back through structured to raw. Agents may never skip a layer.

Around the engine: an MCP interface every agent in the fleet shares — the same protocol reads facts before a task and writes lineage after — plus Atlas, an animated graph visualization that renders the entire brain as nodes and wikilink edges, with ghost nodes marking pages that should exist but don't yet. Gaps are rendered, not hidden.

And an audit-trail design that treats failure as data: sync failures land in an append-only log with the offending commit hash; rerank timeouts are written to a weekly audit file; a doctor pass scores the brain's health and names what's degraded. When a malformed journal file crashed ingestion, the system logged the exact commit and error, marked the source corrupted, and kept serving reads — the failure was loud, attributable and contained, which is the property I actually care about in production memory.

These numbers come from the brain's own health tooling, run live while writing this page. The same tooling is equally blunt about what's weak — link density and sync freshness are scored and failing them is visible — because a memory layer you can't audit is just a cache with confidence.