What's this about?
ConvMem summarizes long documents through a parallel hierarchy, treating a query-prompted LLM as a convolution kernel over segments rather than updating one sequential memory.
What exactly is a “logarithmic tree” here?
Many segments are summarized simultaneously, then summaries are combined in successive levels. The number of dependent steps grows like a tree's height instead of one step per segment.
Why is this worth noting?
The method is training-free and can parallelize across document chunks and semantic channels. Reported gains over baselines still need broader evaluation outside two long-context QA suites.