Adaptive beamforming in time-varying environments depends on the length of the window over which the spatial covariance is estimated. A long window yields stable, low-variance estimates but blurs distinct interference states together, while a short window tracks change but starves the estimator of snapshots. Conventional practice fixes this window through sliding window or forgetting factor heuristics that presuppose a constant timescale of stationarity. This paper instead infers the window window from the data. Under the distortionless constraint, the Capon beamformer is shown, through its generalized sidelobe canceler form, to be a constrained linear predictor, so that minimizing output power over an interval is identical to minimizing a prediction residual. Bellman's principle of optimality then partitions the record into variable-length stationary segments, yielding the Online Segmented Beamformer (OSB). The OSB a causal algorithm that maintains a bank of candidate beamformers and resets its covariance at data-driven change points, with regret growing only logarithmically relative to the best segmentation chosen in hindsight. The selection it performs is a form of content-addressed attention over past snapshots. Experiments in a reverberant moving-source simulation, on the SwellEx-96 shallow-water array, and on a distributed microphone array show that the OSB matches the agility of short windows and the precision of long integration without a fixed prior on the coherence time.