Intuition with Evidence: Adaptive Dual Memory for Multimodal Agents
Authors: Qiuhe Hong, Jingya Wang, Cuzhen Geng, Shenyao Peng, Munan Ning, Yuyang Liu, Yonghong Tian
Subject: Computer Vision and Pattern Recognition (cs.CV)
arXiv: 2610.33019 [cs.CV]
Submitted: 8 October 2026
Overview
As multimodal agents engage in sustained interactions, long-term memory must serve two distinct but complementary functions: fine-grained evidence recall and internalized interpretation for global understanding. Episodic and latent memory each address one of these demands—through explicit historical records and internalized experience, respectively—but neither paradigm alone can accommodate the diverse memory affinities across tasks. This limitation stems from gaps in episodic evidence coverage and integration, as well as from weak factual grounding in latent memory.
Proposed Framework
To bridge this divide, the authors propose a dual-memory framework that unifies:
- A model-native latent memory, functioning as an intuitive prior, and
- A multi-indexed episodic memory, enabling exact evidence reconstruction through retrieval-conditioned joint alignment.
At inference time, the framework adaptively composes these two memories via a training-free routing mechanism. Routing is driven by Jensen-Shannon Divergence (JSD) sensitivity and confidence proxies, which together estimate query-dependent memory affinity.
Results
Extensive evaluations across multimodal benchmarks demonstrate state-of-the-art performance. The framework outperforms single-paradigm baselines overall, achieving a +4.67% improvement over the vanilla episodic baseline. Further analysis highlights episodic memory's strengths in evidence-localized tasks, alongside latent memory's contributions to integrating dispersed information and compensating for imperfect retrieval.
Significance in the 2026 Context
As multimodal agents move toward longer-horizon, sustained interactions—spanning embodied assistants, document-and-video reasoning, and tool-using vision-language systems—robust long-term memory has become a central bottleneck. This work's training-free, adaptively routed dual-memory design aligns with the field's broader shift toward modular, inference-time composable memory architectures that balance interpretability (evidence) with generalization (intuition).
Reference
Qiuhe Hong, Jingya Wang, Cuzhen Geng, Shenyao Peng, Munan Ning, Yuyang Liu, and Yonghong Tian. Intuition with Evidence: Adaptive Dual Memory for Multimodal Agents. arXiv:2610.33019 [cs.CV], 2026. DOI: https://doi.org/10.48550/arXiv.2610.33019
via ArXiv CV