This paper identifies a previously unmeasured stage in multimodal memory systems called delivery: the process of formatting retrieved records for the answering model. While most research focuses on what is stored and retrieved, the authors show via controlled experiments on MemLens that delivery matters more. With the retrieved set fixed to gold evidence, delivering original pixels instead of withholding them raises accuracy by 13.87 points on an 8B backbone, while making retrieval perfect on the same records adds only 2.31 points. The gap persists across larger backbones.
The authors propose DeliverMem, a delivery strategy with three decisions: keep the original visual modality, give each item a readable identity, and state when it was seen. A retrieval-side adapter handles a property delivery cannot supply. Each decision is independently measured against a control that changes only that decision. DeliverMem outperforms the strongest published memory agent on MemLens at all four context lengths and beats the strongest method on DMV-Bench at every setting, all while using a tenth to a seventieth of the input. The results are achieved without any training or modification to stored records.