Researchers from Fudan University and other institutions published a survey on October 14 proposing a two-axis framework for how multimodal agents form, represent and manage memory as they learn over long periods. The paper, submitted to arXiv as 2610.19366, addresses the challenge of retaining useful information from complex, varied data while supporting continuity across sessions and adapting to changing users and environments.

The first axis, Formation and Representation, describes what information is retained and how it is encoded. Examples include modality-specific embeddings (numerical representations of particular data types), textual or structured records, and latent or parametric representations. These forms are intended to support later reasoning and action. The second axis, Memory Management, divides systems into three tiers: T1 Static Memory reuses fixed memory; T2 Dynamic Memory updates memory through a stable policy; and T3 Self-Evolving Memory updates memory through a policy that changes over time.

The survey distinguishes updating stored content from improving the policy that manages it. It synthesizes design principles for balancing fidelity to source information with abstraction, preserving temporal validity and provenance (the record of where information came from), grounding memory use in evidence that can be recovered, and evolving management policies through experience. The authors also cover memory infrastructure, evaluation benchmarks, applications and open challenges for multimodal agents.

Researchers from Fudan University and other institutions published a survey on October 14 proposing a two-axis framework for how multimodal agents form, represent and manage memory as they learn over long periods. The paper, submitted to arXiv as 2610.19366, addresses the challenge of retaining useful information from complex, varied data while supporting continuity across sessions and adapting to changing users and environments.

The first axis, Formation and Representation, describes what information is retained and how it is encoded. Examples include modality-specific embeddings (numerical representations of particular data types), textual or structured records, and latent or parametric representations. These forms are intended to support later reasoning and action. The second axis, Memory Management, divides systems into three tiers: T1 Static Memory reuses fixed memory; T2 Dynamic Memory updates memory through a stable policy; and T3 Self-Evolving Memory updates memory through a policy that changes over time.

The survey distinguishes updating stored content from improving the policy that manages it. It synthesizes design principles for balancing fidelity to source information with abstraction, preserving temporal validity and provenance (the record of where information came from), grounding memory use in evidence that can be recovered, and evolving management policies through experience. The authors also cover memory infrastructure, evaluation benchmarks, applications and open challenges for multimodal agents.