Memory as a Usable Resource: Correlations, Records and Complete Endpoint Accounting

A stored correlation has value only through an operation that can use it. We ask which correlations can fund a free-energy gain, who can read them and what reading, feedback and memory reset cost. The survival functional $\mathcal S=\Delta F-W$ obeys an exact balance retaining mutual information, environmental displacement from equilibrium, control energy and interaction energy. Three models distinguish their roles: Lorentz–Drude dephasing has no backflow at any temperature but a nonzero interaction energy; an autonomous two-qubit recurrence saturates the correlation budget without a pointer record; fresh-memory collisions create records available to several readers. A classical memory used for feedback is worth at most $k_BT I(X{:}M)$, with matched quasistatic feedback approaching the value. Every feedback protocol assumes a posterior and loses at least $k_BT$ times its expected divergence from the true posterior. A wrong readout model can therefore make a record worse than no memory. For a tail-controlled estimate from $n$ references per sign, calibration has leading loss $k_BT(1+g^2/(1-b^2))/(4n)$ and pays through reuse. Within a reset menu with no added purity, positive read–harvest–reset work requires accessible correlated side information. Two partially mixed record ports realize this possibility: one drives matched feedback and the other reduces the reset cost through a reversible parity operation. These results establish the memory–self–introspection route's first functional link: relevant memory permits a gain, a model of one's readout realizes its full value, and calibration builds that model at a calculable cost. A separate finite-history classification bound proves that, for each fixed observation window, a longer finite episode can improve the attainable decision risk.