Don't Scroll Back: Missing-Evidence Memory for Streaming Dialogue Summarization
The paper argues that streaming summaries fail when memory misses the old evidence assumed by the current dialogue window.
Min and Song define the task as summarizing a live window with selective memory from an unbounded prior conversation. Their benchmark separates two checks: whether memory retrieves the missing evidence, and whether the summary uses it. ReMEMBER retrieves against unresolved dependencies, then compresses chunks into evidence-dense memory within a fixed budget. In tests with histories up to 160K tokens, it improves memory recall and gap-resolution completeness over baseline memory methods. HF Daily Papers' note
Min and Song define the task as summarizing a live window with selective memory from an unbounded prior conversation. Their benchmark separates two checks: whether memory retrieves the missing evidence, and whether the summary uses it. ReMEMBER retrieves against unresolved dependencies, then compresses chunks into evidence-dense memory within a fixed budget. In tests with histories up to 160K tokens, it improves memory recall and gap-resolution completeness over baseline memory methods. HF Daily Papers' note
score 5