echo-memory-train

name	echo-memory-train
description	Run Echo-Memory memory-baseline and context training recipes on Wan 2.1 1.3B. Use when training spatial/SSM/compression/context rows, editing train/*.sh launchers, or configuring DATASET_BASE_PATH for static or dynamic pools.

Echo-Memory training

Required env (repo root)

export WAN_BASE_MODEL=/path/to/Wan2.1-T2V-1.3B
export DATASET_BASE_PATH=data/Context-as-Memory-Dataset   # static in-domain pool
export PYTHONPATH=$PWD:${PYTHONPATH:-}
export OUTPUT_BASE_ROOT=$PWD/outputs

• Static in-domain pool — default data/Context-as-Memory-Dataset; see doc/dataset_preprocessing.md
• Dynamic training pool — e.g. data/dynamic-memory-dataset; see doc/dynamic_dataset_preprocessing.md

Entry scripts

Family	Directory	Example
Spatial / SSM / compression	train/memory_baselines_basic/	run_spatial_memory_baseline.sh, run_ablation_block_wise_ssm_two_chunk.sh
Context K=1/5/20	train/context_learning/	run_pre_qkv_ctx1.sh, run_pre_qkv_ctx20.sh

Run from repository root: bash train/memory_baselines_basic/run_spatial_memory_baseline.sh

Shared env: train/_shared/common_env_memory.sh — no private paths baked in.

Agent checklist

1. Confirm which memory row (paper family + row id) the user wants.
2. Set DATASET_BASE_PATH to the correct training pool name in docs (Echo terms only in public markdown).
3. Prefer editing existing run_*.sh patterns over new one-off Python entrypoints.
4. Outputs go to outputs/ unless OUTPUT_BASE_ROOT is set.
5. Do not commit data/, outputs/, checkpoints, or machine-local paths.

Docs

• train/README.md — launcher index
• train/memory_baselines_basic/README.md — ablation set