Finite-dimensional Koopman models are increasingly used in predictive control, but their accuracy remains limited by truncation errors and offline identification data. Adaptive Koopman predictive control promises to address this limitation, but typically requires closed-loop persistent excitation (PE). Since this is difficult to achieve in practice, in this paper we develop an adaptive generalized Koopman (GeKo) predictive controller that does not require PE. To this end, we formulate a recursive matrix regression in fixed lifted coordinates that extends the Kalman filter with back-to-the-prior forgetting to matrix-valued operators and uses a matrix normal invariant prior, centered at the operator identified offline. For any regressor sequence, and thus under any feedback law, we prove that the covariance remains between two multiples of the prior covariance and that the estimation error is input-to-state stable. On an academic example, a differential-drive robot and a power system with plant changes and little excitation, the invariant-prior estimator yields the lowest overall tracking error among the compared estimators and prevents covariance windup.