EM^2Mem: Event-Centric Multimodal Memory for Large Language Models
EM^2Mem stores long-video evidence around events, not loose captions or frames.
The paper says existing multimodal memory systems leave models to reconnect time, modality, and attribution at query time. EM^2Mem instead builds event-indexed memory cells that tie together records, temporal context, graph relations, facts, and provenance. On three long-video QA benchmarks, it reports gains of 2.0, 2.4, and 3.7 accuracy points over the strongest memory baseline, plus higher event-level Top-5 evidence recall. It also reports 4.67x lower per-query latency and 63.66% fewer inference tokens. HF Daily Papers' note
The paper says existing multimodal memory systems leave models to reconnect time, modality, and attribution at query time. EM^2Mem instead builds event-indexed memory cells that tie together records, temporal context, graph relations, facts, and provenance. On three long-video QA benchmarks, it reports gains of 2.0, 2.4, and 3.7 accuracy points over the strongest memory baseline, plus higher event-level Top-5 evidence recall. It also reports 4.67x lower per-query latency and 63.66% fewer inference tokens. HF Daily Papers' note
score 4