---
license: cc-by-4.0
base_model: Wan-AI/Wan2.1-T2V-1.3B
pipeline_tag: text-to-video
tags:
  - world-model
  - video-generation
  - action-conditioned
  - memory
  - ablation
  - research-artifact
---

# echo — Echo-Memory ablation arm (`ucpe` + two-chunk memory)

> **This is an unofficial derivative artifact, not the official Echo-Memory release.**
> It contains one ablation arm re-trained from the Echo-Memory codebase, produced for
> internal study. For the official code, paper and checkpoints, use the links below.

## Upstream / attribution

| | |
|---|---|
| Original work | **Echo-Memory: A Controlled Study of Memory in Action World Models** |
| Authors | Echo Team @ Joy Future Academy, JD |
| Paper | [arXiv:2606.09803](https://arxiv.org/abs/2606.09803) |
| Official code | https://github.com/Echo-Team-Joy-Future-Academy-JD/Echo-Memory |
| Official checkpoints | https://huggingface.co/Echo-Team/Echo-Memory |
| Code license | CC BY 4.0 |
| Base model | [Wan2.1-T2V-1.3B](https://huggingface.co/Wan-AI/Wan2.1-T2V-1.3B) (Apache 2.0) |

The Echo-Memory code included here is redistributed under CC BY 4.0 with attribution to
the Echo Team. The weights are derived from Wan2.1-T2V-1.3B, which is Apache 2.0; that
license and its notices continue to apply to the derived weights.

## What this repository contains

```
code/                       Echo-Memory codebase as used for this run
checkpoints/
  Step-10000.safetensors    3.5 GB
  Step-20000.safetensors    3.5 GB
  Step-30000.safetensors    3.5 GB   (final, max_train_steps = 30000)
```

## Training configuration

Single arm of a memory-mechanism ablation, run name
`memory_baselines_basic_abl_ucpe_memory_two_chunk`.

- Base weights: `Wan2.1-T2V-1.3B` (DiT + `umt5-xxl` text encoder + `Wan2.1_VAE`)
- Trainable: `dit` only, `--remove_prefix_in_ckpt pipe.dit.`
- Memory: `--enable_context_memory --context_source replay`
  `--prev_chunk_frames 81 --context_memory_frames 5`
- Mechanism under test: `--use_cgla_memory --cgla_mechanism ucpe`
  `--cgla_every_n_blocks 4 --cgla_aux_loss_weight 0.01`
- Also enabled: `--use_moc --moc_temperature 1.0`, `--train_cam_pose`,
  `--add_action_attn`, `--use_rt_relative`, `--use_anchor_frame`, `--cfg_target_only`
- Resolution / length: 352 x 640, 81 frames
- Optimisation: lr 2e-5, 1 epoch, batch 1/device, grad-accum 1,
  `--timestep_shift 15`, `--context_drop_prob 0.1`, 30,000 steps
- Hardware: 1x NVIDIA H20

## Data — not redistributed here

Training used the **Context-as-Memory** dataset (Unreal-Engine rendered environments,
~331 GB of frames). It is **not** included in this repository: it is third-party data
and the copy used here carried no redistribution licence. Obtain it from its original
source.

## Status and caveats

- This is **one arm of an ablation sweep**, uploaded as a research artifact.
  Its results had not been analysed at the time of upload — no quality claim is made,
  and it should not be read as the best or recommended configuration.
- It is **not** a reproduction of, or a replacement for, the official Echo-Memory
  checkpoints. Comparisons against the paper's numbers are not valid without matching
  the official evaluation protocol.

## Citation

Please cite the original work:

```bibtex
@article{echomemory2026,
  title  = {Echo-Memory: A Controlled Study of Memory in Action World Models},
  author = {Echo Team, Joy Future Academy, JD},
  journal= {arXiv preprint arXiv:2606.09803},
  year   = {2026}
}
```