Lifelong agents are designed to keep learning after deployment, accumulating experience and reusing it to improve. But that memory is not free: in practice, agents operate under tight budgets for generation, interaction, and memory. The new arXiv paper "Remembering Has a Price" takes these joint constraints as its starting point.
Rather than treating memory as an unlimited store, the authors frame remembering as a resource-allocation problem. An agent must decide which experiences are worth keeping when storing them competes with the cost of generating and interacting. The abstract sets up this trade-off and reports that under such stringent joint constraints, the authors find ... the listing cuts off before the full finding is stated.
The available text does not specify the method or the detailed results. What is clear is the paper's central premise: memory is not a neutral archive but a costed resource, and lifelong learning under real-world limits requires deciding what is worth remembering.