For decades, the community has relied on a familiar methodology: fix a heap size, run a benchmark, and report throughput, pause times, or overall runtime. This approach has driven remarkable progress, but it also encodes a particular view of memory management—one where heap size is static, trade-offs are implicit, and performance is treated as a property of a single configuration rather than a dynamic system. This talk argues that this evaluation model has reached its limits. Drawing on a series of recent studies, I will show that many of the behaviours we attribute to garbage collection are in fact artefacts of how we measure it. Performance varies non-linearly across heap sizes; collectors routinely perform substantial work on objects that are already dead or about to become unreachable; and decisions about when and where to act—scheduling work, adjusting heap size, or placing objects in memory—can dominate outcomes. When examined across these dimensions, memory management reveals itself not as a static optimisation problem, but as a complex, interacting system spanning time, space, and hardware locality.

These observations point toward a different perspective: garbage collection should be understood as a feedback-driven control problem. Instead of fixing resources and observing behaviour, we can invert the relationship—treating memory, CPU overhead, and locality as dynamically adjustable variables, guided by explicit performance goals. Early examples of this approach already suggest that systems can continuously trade memory for computation, defer or advance work based on runtime conditions, and reorganise memory layouts based on observed access patterns.

The goal of this keynote is not to present a single new mechanism, but to challenge how we reason about memory management itself. By moving beyond static benchmarks to policy-driven, adaptive evaluation, we open up a richer design space—one where garbage collectors are not just mechanisms for reclaiming memory, but active participants in shaping program performance.