NativeMEM

Paper | Code | Project Page

NativeMEM equips a pretrained vision-language-action (VLA) policy with long-term, real-time-updated visual memory. Its Native Memory Compression scheme repurposes the VLA's own vision encoder to compress each historical frame-view observation into a single native memory token, then appends those tokens to the VLA's original input sequence.

Checkpoints

Checkpoint	Path
Stage 1 pretrained memory tokenizer	checkpoints/mem_tokenizer_pretrain/pretrain/49999/params
Stage 2 task-specific NativeMEM policies	checkpoints/nativemem_pi05/<task>/19999

Each Stage 2 checkpoint contains params and assets; the normalization statistics in assets are reloaded at evaluation.

Usage

Download into the root of the code repository, which uses the same checkpoints/ layout:

hf download Ziyeeee/NativeMEM --include "checkpoints/mem_tokenizer_pretrain/pretrain/*" --local-dir .
hf download Ziyeeee/NativeMEM --include "checkpoints/nativemem_pi05/*" --local-dir .

See the code repository for training and data preparation.