Survey Maps 297 Multimodal Agent Memory Designs on Two Axes

A large multi-author survey on Hugging Face Papers catalogues 297 multimodal agent memory architectures and 82 benchmark records, and proposes a two-axis framework to make sense of them. Axis I covers what information is retained and how it is encoded; Axis II sorts systems into three tiers of memory management.

The tier labels are the usable part. T1 Static Memory "stores and reuses fixed memory without modification." T2 Dynamic Memory "updates memory state through stable policy" via operations like appending, revising, merging, consolidating, invalidating, and forgetting. T3 Self-Evolving Memory goes further and updates "both memory state AND the policy governing memory management." The paper's reading of the field is that T2 is widespread while deployment-time T3 systems remain scarce.

On representation, the authors group designs into source- and modality-specific encodings, cross-represented explicit memory such as perception-to-text records and scene graphs, internalized latent and parametric memory including hidden states, memory tokens, KV caches, and LoRA adapters, and hybrid combinations with linked evidence. They describe hybrid architectures as prevalent in advanced systems.

The survey tracks activity from 2023 through 2026 and reports that growth is steepest in embodied and video memory, moderate in image and GUI work, and comparatively sparse in audio and speech. It lands alongside this week's other agent-memory and VLA pieces in our agents coverage, including NVIDIA's TacitVLA latency result from the day before.

The authors frame open directions as bounded lifelong memory, governed forgetting, lifecycle-aware evaluation, cross-representation consistency, and deployment-time policy evolution.