Code agents increasingly operate on real repositories that change over time. To work efficiently, they can rely on memory systems that store and retrieve earlier experience. But that experience is diverse and constantly evolving, and it remains unclear how memory systems should organize it for effective reuse.
The paper introduces MemDev-Bench, a benchmark designed to test memory systems over real repository histories. Its central question is whether memory actually helps code agents, and it provides a structured way to measure that.
The abstract emphasizes the open problem of organizing varied and shifting experience, suggesting that MemDev-Bench is a step toward understanding when and how memory systems benefit code agents in realistic settings.