The Cowork Memory Hub

Challenge

Running multiple brands and projects through an AI assistant surfaced a real, non-obvious problem: the same business facts were being held across as many as seven disconnected context layers — chat memory, project instructions, account settings, preference stores, auto-regenerating project memory — sometimes in four mutually contradictory states at once, with no way to tell which was current.

Approach

Designed and built a two-tier memory system — a working file plus a structured knowledge directory — as the single source of truth, governed by one principle: layers that carry facts go stale, layers that carry routing don’t. Built two automated audit agents that diff every layer against the hub, flag conflicts, and generate ready-to-use corrections, backed by a documented, reusable protocol.

Result

Verified in production, not just designed: the first live audit run caught the protocol’s own assumptions being wrong twice and surfaced nine real conflicts plus a fabricated-figure risk. In active, daily use across every brand and project this site draws from — including this case study catalog itself.

Lesson: the self-referential nature is the strongest proof point, not a caveat — it’s a working system solving a problem most AI users don’t yet know they have, with its own audit trail as evidence.