Reference Frames Selected by History: The Capability and Cost of a Bounded Agent's Self-Frame

How does a memory begin to organize the system that carries it? We study the step in which retained history selects what the system will read next. A finite two-channel task first exhibits the mechanism: training records form a two-bit signed-coordinate code whose one-read risk beats every history-independent direction, even with an exact coefficient supplied to that comparator. A new control allocates the same two retained bits to the same history-selected axis and an unsigned confidence flag. It has identical foreground alignment but no advantage over predicting zero. We characterize the optimal encoder for the fixed four-action decoder, without claiming optimality among all two-bit representations. General capability and information bounds then explain what such access requires. The frame's own records leave omitted directions unidentifiable; explicit drift and calibration models show when outside evidence pays. A finite quantum reference provides a separate physical realization. The subject-formation proposal is that learned relations organize the same system's future access, protection and revision. The models establish components of that proposal, not a statistical distinction from general learning or a derivation of experience. Main and Appendices PDFs are supplied separately, with full sources and scientific supplements in the accompanying ZIP.