A new survey on arXiv examines the shift from simply extending LLM context windows to building selective memory systems. The authors note that while long-context capability is now essential for complex tasks, continuously expanding the window does not eliminate the underlying pressures on computation, storage, and information selection.
The survey focuses on memory-centric compression, an approach that treats context not as an ever-growing input but as a memory to be curated. Instead of keeping everything, these methods aim to compress and retain only what matters, which could reduce resource demands while maintaining task performance.
Because this is a survey, it consolidates existing work rather than presenting new experiments. The abstract does not specify which methods are most effective, but it frames memory-centric compression as a promising direction for making long-context LLMs more practical.