Echo-Memory — Wan 2.1 1.3B memory baseline checkpoints

Paper-aligned epoch-0 fine-tunes for Echo-Memory.

Paper | Project Page | GitHub

Backbone: Wan-AI/Wan2.1-T2V-1.3B
Training: static in-domain pool · 1 epoch · 30,000 steps · 640×352 · 81-frame chunks
Layout: {row_id}/epoch-0.safetensors

Checkpoint index

Family	Paper row	HF path	Steps
Raw context	Context K=1	context_k1/epoch-0.safetensors	30,000
Raw context	Context K=20	context_k20/epoch-0.safetensors	30,000
Spatial	Spatial Memory	spatial_mem/epoch-0.safetensors	30,000
State-space	Block-wise SSM	block_wise_ssm_two_chunk/epoch-0.safetensors	30,000
State-space	Legacy Hybrid (VideoSSM)	videossm_hybrid/epoch-0.safetensors	30,000
Spatial	concat text (ablation)	spatial_concat_text_two_chunk/epoch-0.safetensors	30,000
Spatial	inject none (ablation)	spatial_inject_none_two_chunk/epoch-0.safetensors	30,000
Spatial	cross-attn t32 (ablation)	spatial_cross_attn_readout_t32_g4_two_chunk/epoch-0.safetensors	30,000
State-space	SSM ctx1 / every4 / hint21	ssm_ablation_ctx1_every4_hint21/epoch-0.safetensors	30,000
State-space	SSM ctx5 / every1 / hint21	ssm_ablation_ctx5_every1_hint21/epoch-0.safetensors	30,000
State-space	SSM ctx5 / every4 / hint81	ssm_ablation_ctx5_every4_hint81/epoch-0.safetensors	30,000

Context K=5 and FramePack compression rows are not yet released as epoch-0 weights.

Download

pip install -U "huggingface_hub[cli]"

# one row
huggingface-cli download Echo-Team/Echo-Memory context_k1/epoch-0.safetensors --local-dir ./ckpts

# all rows
huggingface-cli download Echo-Team/Echo-Memory --local-dir ./ckpts

Keep the row subdirectory in the local path (e.g. ./ckpts/spatial_mem/epoch-0.safetensors).

Use with Echo-Memory code

Clone Echo-Memory, install the environment, then:

export WAN_BASE_MODEL=/path/to/Wan2.1-T2V-1.3B
export DATASET_BASE_PATH=data/Context-as-Memory-Dataset
export PYTHONPATH=$PWD:${PYTHONPATH:-}
export CKPT=./ckpts/spatial_mem/epoch-0.safetensors

# in-domain replay + revisit
bash eval/v2/run_static_consistency_loop_and_revisit.sh
bash eval/v2/run_basic_replay_gt.sh

# open-domain revisit (first frames in repo)
PHASE=stage1 OOD_DIR=assets/opendomain_revisit \
  bash eval/v2/revisit_suite/run_one_click_revisit_eval.sh

Memory runtime flags are inferred from the checkpoint path via env/memory_baseline_runtime.py — use the HF folder names above.

Citation

If you use this repository or the Echo-Memory paper, please cite:

@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},
  month={jun},
  eprint={2606.09803},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2606.09803}
}