Position: Context Length Is a System Limit, Not a Measure of Memory

Maximum context length tells us how much text a system can accept, not what it retains, retrieves, updates, or preserves beyond the active context. We argue that it should be treated as a system limit rather than the headline measure of memory. We propose a three-part memory report. A *memory specification* $\mathcal{M}=(R,A,U,P)$ describes the mechanism: what is stored, how it is accessed, how it changes, and when it is reset. A *memory profile* measures performance on distinct memory tasks. A *resource report* states the limits and costs under which those results were obtained. Together, the three parts connect mechanism, behavior, and cost, giving a practical reporting template for model cards, technical reports, and system papers. We apply the framework to recent systems and use a 128K-token project example to show why recent-window, retrieval, compressed-state, and full-context systems can rank differently. The proposal is testable: the complete report should predict long-horizon performance better than supported length alone.