Intelligent Memory: A Theory of Minimum Sufficient Memory under Lifecycle Energy Constraints

Persistent artificial-intelligence systems must decide not only what to retrieve from the past, but what to store, transform, maintain, expose to a reader model, and eventually forget. These operations have different costs over the lifetime of a memory, yet contemporary systems are commonly optimized for accuracy, token count, latency, or a local memory budget rather than physical lifecycle energy. We introduce Intelligent Memory (IM), a formal framework in which memory management is posed as a constrained control problem: choose the least-energy memory policy that keeps robust decision distortion below a declared tolerance. The framework derives five operational principles: minimum sufficient transformation, counterfactual memory value, event-triggered memory dynamics, minimum-energy sufficient context, and resource-constrained selective forgetting. Evidence traceability is maintained independently of memory activation, while storage is treated as a lifecycle quantity rather than a peak-size statistic. The theory is implementation-agnostic: it specifies admissibility conditions for memory actions without prescribing a particular storage layout, deletion mechanism, retrieval scheduler, or hardware-control loop. A controlled proof-of-concept evaluates the routing logic in a synthetic counterfactual environment rather than on a public leaderboard. On held-out controlled regimes, an IM controller reaches 87.40% expected query accuracy and 99.85% event-level sufficiency, with approximately 3% lower modeled energy per expected correct answer than an always-Hybrid policy and a modeled lifecycle cost within approximately 3% of a resource oracle. The same controller uses 27.1% less active storage than always-Hybrid before structural forgetting; a conservative forgetting policy reduces lifecycle storage by 22.9% in a separate stress test. These energy values are synthetic cost units, not joules. We therefore make no claim of state-of-the-art or physical-energy superiority. Instead, IM provides a falsifiable theory and a pre-specified hardware validation protocol for testing whether sufficient intelligence can be maintained with less memory work.