Echo-Memory: A Controlled Study of Memory in Action World Models

arXiv 2026 62.7 method

TLDR

A controlled study comparing memory mechanisms in action-conditioned world models, isolating capacity, compression, read-out, and recurrence axes.

Reasoning

Strengths: Clear isolation of memory design factors under a fixed backbone, with a three-branch evaluation protocol revealing that replay quality is insufficient for world memory. Weaknesses: The abstract cuts off mid-sentence, leaving some findings incomplete; no explicit mention of real-world datasets or benchmarks beyond the evaluation protocol.

Read-first score

Read-first score 62.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 45.

Recency 6%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Citation impact 18%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.94977106

Methodology quality 18%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=baseline,evaluation,metric,result

Topical relevance 29%

64.3

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FoundationFrontierBridgeMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 433.

Keyword Scores

world model

video world model

generative world model

interactive world model

world dynamics prediction

world simulator

model-based reinforcement learning world model

Deep Analysis

Innovations

Controlled study isolating memory mechanisms in action world models by fixing all other components (backbone, optimizer, camera-action representation, sampler, evaluation pipeline).
Three-branch evaluation protocol (replay quality, in-domain loop revisit, open-domain return probes) that reveals disagreement among metrics, showing replay fidelity is not a sufficient proxy for world memory.
Key findings: raw context as a strong capacity baseline, compactness not a free substitute for capacity, and block-wise state-space recurrence as the strongest open-domain return mechanism.

Methodology

Echo-Memory fixes the action-to-video interface and varies only how history is stored and read by the generator. Under a shared video diffusion backbone, optimizer, camera-action representation, sampler, and evaluation pipeline, it compares raw context, compression-based memory, spatial summaries with different read-out paths, and state-space recurrence, separating four axes: capacity, compression, read-out, and recurrence.

Key Results

Raw context improves open-domain return far more than replay metrics; aggressive spatial and hybrid-compression memories lose salient evidence needed for return; block-wise state-space recurrence is the strongest open-domain return mechanism in the matrix.

Limitations

Limited to a single video diffusion backbone, so generalizability to other architectures is not tested.
Evaluation protocol may not capture all aspects of memory in real-world or diverse scenarios.
The study does not explore all possible memory mechanisms (e.g., attention-based or learned retrieval).

Resources

Paper