Echo-Memory — Wan 2.1 1.3B memory checkpoints

Validated 30,000-step fine-tunes for Echo-Memory, built on Wan-AI/Wan2.1-T2V-1.3B.

Paper | Project page | Training and inference code

Available checkpoints

Family	Checkpoint	Steps	SHA256
Raw context	context_k1/epoch-0.safetensors	30,000	See LFS metadata
Compression	framepack_len_r8/epoch-0.safetensors	30,000	dd57625506a2c68c402dc05de8f3c6fc5f5376fcac77d7fb6f26eb6ace1d74bf
State-space	block_wise_ssm_causal_v2/epoch-0.safetensors	30,000	0dd90ea3f3423644f4d68c6d1185d7d717d328f5922ee21f745fc85abe9a01a9

MoC, geometry-grounded spatial memory, legacy SSM, and VideoSSM weights are not part of this validated release.

Download

huggingface-cli download Echo-Team/Echo-Memory \
  block_wise_ssm_causal_v2/epoch-0.safetensors \
  framepack_len_r8/epoch-0.safetensors \
  --local-dir ./ckpts

Keep the row folder in the local path. Echo-Memory auto-detects the profile and restores the trained context order, FramePack strategy, and camera-RT continuation.

Multi-chunk inference

export WAN_BASE_MODEL=/path/to/Wan2.1-T2V-1.3B
export CONTEXT_IMAGE=/path/to/first_frame.png
export ACTION_PATH=/path/to/81_frame_camera_action.json

CKPT=./ckpts/block_wise_ssm_causal_v2/epoch-0.safetensors NUM_CHUNKS=4 \
  bash inference/memory_baselines_basic/run_infer_block_wise_ssm_causal_v2.sh

CKPT=./ckpts/framepack_len_r8/epoch-0.safetensors NUM_CHUNKS=4 \
  bash inference/memory_baselines_basic/run_infer_framepack_len_r8.sh

Citation

@article{king2026echomemory,
  title={Echo-Memory: A Controlled Study of Memory in Action World Models},
  author={King, Wayne and Xue, Zeyue and Bian, Yuxuan and Huang, Jie and Li, Haoran and Li, Yaowei and Su, Yaofeng and Li, Yuming and Wang, Haoyu and Zhang, Shiyi and Zhang, Songchun and Niu, Yuwei and Xu, Sihan and Zhuang, Junhao and Huang, Haoyang and Duan, Nan},
  journal={arXiv preprint arXiv:2606.09803},
  year={2026}
}