WorldFoundry
English中文

Echo-Memory

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

JD Echo-Memory identity permanently pinned to context_k1 (K=1). Wan2.1-1.3B. CC-BY-4.0. 2026-07-20 A100 clip is historical — native CUDA artifact pending.

Edit this pageGitHubMarkdown

Echo-Memory Context K=1 is one identity in JD's Echo-Memory family: a Wan2.1-T2V-1.3B interactive world that differs from siblings only in memory. This card is permanently bound to context_k1/epoch-0.safetensors — each new chunk conditions on K=1 context frame. You do not pick a memory method at runtime. Pipeline: EchoMemoryContextK1Pipeline. Aliases echo-context-k1 and echo-memory resolve here.

Official paper teaser: Abstract teaser and workflow of Echo-Memory. Given a text description, historical observations, and the camera/action state, an action world model must generate chunk-wise video while carrying memory across revisits..
Official paper teaser: Abstract teaser and workflow of Echo-Memory. Given a text description, historical observations, and the camera/action state, an action world model must generate chunk-wise video while carrying memory across revisits..

Paper: arXiv 2606.09803. Project: echo-team-joy-future-academy-jd.github.io/Echo-Memory. Code: Echo-Team-Joy-Future-Academy-JD/Echo-Memory. Weights: Echo-Team/Echo-Memory (cc-by-4.0) plus Wan-AI/Wan2.1-T2V-1.3B (apache-2.0).

A 2026-07-20 A100 artifact predates the NativeDiffusionPipeline cutover and is historical only. Fresh CUDA artifact, official sample parity, and bench scoring are pending. Visual quality is research-preview (1.3B).

Build your run

Pipeline
EchoMemoryContextK1Pipelineworldfoundry.pipelines.echo_memory.pipeline_echo_memory
TaskDefault default
Environmentworldfoundry-unified-cu128
DeviceCUDA 12.8
Runnerworldfoundry.pipeline
Loadingfrom_pretrained
Invocationnative_or_standard_invocation
worldfoundry-eval run \
  echo-memory-context-k1 \
  --pipeline.task-profile default \
  --json

What it is for

Echo memory-context K1 slots, not the Infinity long-video card.

Launch worldfoundry-eval run echo-memory-context-k1.

Score on MemoBench — purpose-built for evaluating long-horizon memory in interactive world models.

Runtime contract

Required: prompt and image; optional: video; actions: camera_trajectory; artifacts: generated_world_video and echo-memory-context-k1.mp4.

KnobDefault
Size640×352
Frames81
FPS15
Chunks2
Context frames1
Sample steps50
Sample shift5.0
Camera trajectoryz*80 (stationary)
Camera translation step0.00125 per frame (w*80 moves +Y 0.1 per chunk)
Camera rotation step0.5625° per frame (left*80 turns 45° per chunk)
FinetuneEcho-Team/Echo-Memory (cc-by-4.0)
BackboneWan-AI/Wan2.1-T2V-1.3B (apache-2.0)

Pass camera_trajectory="w*80" to request the published forward action scale, or "left*80" for a 45° turn. A camera move mentioned only in the prompt does not replace the default stationary action. camera_translation_step and camera_rotation_step_degrees can override the per-frame distances.

Run in WorldFoundry

Status: Runner parity pending. Unified environment worldfoundry-unified-cu128 (Python 3.11, CUDA 12.8). Stage Echo-Team/Echo-Memory, Wan-AI/Wan2.1-T2V-1.3B, and Echo-Team/Echo-Memory among others locally — missing files fail at load, nothing is downloaded. Run worldfoundry-eval run echo-memory-context-k1. License: CC-BY-4.0 and apache-2.0. Catalog VRAM is unset.

worldfoundry/base_models/diffusion_model/models/networks/echo_memory holds checkpoint-specific network blocks only. It is not a CLI. Memory lifecycle, model assembly, and the runtime contract stay in the shared diffusion infrastructure. How to add a new immutable recipe is on the inference guide. Checkpoint staging is on local assets.