When working memory works: Selective engagement as a tradeoff between cost and flexibility

Working memory (WM) has long been viewed as a store that must be engaged to maintain task-relevant information. Yet growing evidence shows that the brain often avoids activating WM, relying instead on perception or long-term memory (LTM). What determines whether and when WM is engaged? Rather than assuming obligatory maintenance, this review documents systematic patterns of WM avoidance and reactivation across perceptual, mnemonic, and control domains. I propose that these patterns reflect a tradeoff between metabolic and cognitive costs and behavioral flexibility. When perception or LTM can support behavior with low effort and stable content–context bindings, WM is bypassed through external or internal offloading. WM engagement emerges when behavior requires verifying or updating the match between memory content and the current context. From this perspective, WM activation is not the default state but a strategic investment: recruited only when low-cost perceptual or LTM reliance cannot guarantee correct content–context binding. Integrating behavioral, electrophysiological, and computational approaches, this framework unifies disparate findings under a single principle: WM engagement reflects a conditional control policy governing not only what is stored, but when active maintenance is recruited.