Gumclaw is an AI company operating system built around a shared taste file. It turns founder corrections, customer-support rules, and brand standards into checkable criteria that agents can apply before customer-facing work goes live.
This is more than agent memory. The important object is not a log of past answers, but a durable standard for deciding whether new output is good enough to reach a user.
When a founder or manager corrects an agent result, the correction becomes more than a one-off edit. It becomes a checkable rule. The next support reply, landing copy, policy answer, or product text can be evaluated against that layer before it reaches a user.
That changes the role of “company context”. It stops being a folder of notes and becomes a test suite for behavior.
New Runtime read: company memory becomes valuable when it can reject bad output, not merely recall old facts.