Multimodal Agent MemorySubmit a paper

Embodied Memory

100 systems and 23 benchmarks whose primary modality is embodied & 3D.

ImageVideo & streamingAudio & speechGUI & computer use

How they represent memory

How T2 systems change it

Representation × management tier

T1 Static Memory 17 · 17% T2 Dynamic Memory 79 · 79% T3 Self-Evolving Memory 4 · 4% Total
Source- and modality-specific 4 8 – 12
Cross-represented explicit 3 23 2 28
Internalized latent & parametric 3 9 – 12
Hybrid 7 39 2 48

Each system counted once, at its highest tier. Circle area shows the value. Hover to compare with chance; click to list the systems.

What its benchmarks test

Directly evaluated, of 23 benchmarks in the survey's Table 2
  • Cross-modal evidenceCE9
  • Long-range evidenceLE23
  • State changesSC14
  • Conflicting memoriesCF0
  • Unanswerable queriesUQ1
  • Memory-grounded actionMA19
  • Spatial memorySM19
  • Object-state trackingOS17
  • Task-state trackingTS14

Benchmarks

All 100 systems