Multimodal models and agents are increasingly being built to learn from and interact with the real world by integrating information across multiple modalities. Yet as these agents move toward long-horizon tasks and lifelong settings, continually acquiring useful information from complex, heterogeneous multimodal data remains a significant challenge.
To address this, a new survey titled "Seeing, Maintaining, and Learning: The Evolution of Multimodal Agent Memory" introduces a two-axis analytical framework for multimodal agent memory. The first axis, Formation and Representation, examines what information is retained and how it is encoded—ranging from modality-specific embeddings to textual and structured records, as well as latent and parametric representations—to support later reasoning and action.
The second axis, Memory Management, distinguishes three tiers: T1 Static Memory, which reuses fixed memory; T2 Dynamic Memory, which updates memory through a stable policy; and T3 Self-Evolving Memory, which updates memory through an evolving policy. The authors stress that updating memory content and improving the policy that manages memory are distinct capabilities.
Drawing on this framework, the survey synthesizes recurring design patterns and derives principles for balancing source fidelity with abstraction, maintaining temporal validity and provenance, grounding memory use in recoverable evidence, and evolving memory policies over time. It also covers memory infrastructure, evaluation benchmarks, applications, and open challenges.
The authors position the work as a guide for researchers and practitioners building multimodal agents, and invite feedback, discussion, and suggestions for relevant work.