Multimodal Agent MemorySubmit a paper

Seeing, Maintaining, and LearningThe Evolution of Multimodal Agent Memory

Memory · Observatory

memory systems
297
benchmarks
82
modalities
5
Browse all systems

01 /System scan

Image 1/14
VISTA

VISTA: A Visual Harness for Reasoning in an Interactive World

  1. Retained source evidence
  2. Hierarchical multi-tier store
  3. Planner / tool-assisted retrieval

T2 Dynamic Memory2026-10-01

Latest systems

  1. 2026-10 VISTA
  2. 2026-08 VoiceMem
  3. 2026-08 WnW (Waxing-and-Waning KV Cache)
  4. 2026-08 FinAcumen (Financial Memory + Financial Tools harness)
  5. 2026-08 Stateful Streaming ASR (State Carry-over / State Rollback)
  6. 2026-08 ObsMem
  7. 2026-08 GeoForge
  8. 2026-08 StreamFlow

02 /Memory sphereone arm per modality, time along the arm

03 /Representation spectrumsystems per stored form

Source100Encodings78Text78Structured159Latent39Parametric25

04 /Management ladderwhat can change

  1. T1 Static Memory 76

    Built once, then reused

  2. T2 Dynamic Memory 204

    Memory updates; the policy stays stable

  3. T3 Self-Evolving Memory 17

    Feedback evolves the policy

05 /Paper riverevery system by publication date

Jan 2027 · 297 systems
Image 64
Video & streaming 84
Audio & speech 8
GUI & computer use 41
Embodied & 3D 100
2023202420252026

06 /Frontier spotlightthe newest self-evolving systems

T3 Self-Evolving Memory · 2026-08-11

GeoForge

GeoForge: Non-Parametric Self-Evolving Agents for Earth-Observation Reasoning

Evolves its retrieval policy · stores memory as structured records & graphs, text records & summaries

Image

Find your way in

Start from the question you have. Each view uses the same corpus.

Or jump to a modality: ImageVideo & streamingAudio & speechGUI & computer useEmbodied & 3D

What the corpus shows

  1. 68.7% of systems update memory through a stable management policy (T2).

    T2 Dynamic Memory is the norm. Only 17 systems let feedback change the policy itself.

    See the tiers
  2. 46.8% of systems are hybrids.

    Most pair explicit records with source-level evidence or modality-specific encodings.

    See hybrid patterns
  3. 17 of 17 self-evolving systems keep explicit memory.

    Policies evolve where memory can be read and edited; none of the source-only or latent/parametric-only systems do.

    See the matrix
  4. 7 of 82 benchmarks directly test conflicting memories.

    Evaluation favors finding old evidence over resolving conflicting or outdated memories.

    See benchmark coverage

Cite the survey

Guo et al. (2026). Seeing, Maintaining, and Learning: The Evolution of Multimodal Agent Memory. arXiv:2610.19366.