Multimodal misinformation detection has advanced from text-only classification toward joint processing of language, imagery, and social metadata. Yet much of the literature still treats adaptation as an internal fusion problem rather than a systems-level routing problem, leaving open the question of when multimodal escalation is warranted and how expert usage can be reduced without materially weakening predictive quality. This paper introduces MemANS, an event-driven, in-memory Agentic Name Search architecture for multimodal misinformation detection. MemANS treats inference as a resolution process over named experts: a default text–metadata resolver handles ordinary cases, while a specialist is invoked only when cross-modal disagreement and confidence conditions jointly indicate that additional reasoning is justified. The empirical study is conducted on a balanced, fully image-available Fakeddit benchmark comprising 5400 training instances, 840 validation instances, and 840 test instances across six misinformation classes. On the final test set, MemANS achieves 0.6182 macro-F1, 0.6167 accuracy, 0.5405 MCC, and 0.8726 weighted one-vs-rest AUC, clearly surpassing the always-on fusion baseline (0.5732 macro-F1) with a statistically supported gain of 0.0454 (p < 0.001) while using only 1.7202 average experts versus 2.0000. The results indicate that Agentic Name Search is a viable and practically meaningful multimedia-systems abstraction for adaptive multimodal reasoning.