arXiv preprint · 2026

Remember Smarter

Visual History Compressor and Hyperbolic Experience Space for Robotic Memory

Dai Zhou, Jiexi Yan, Tong Li, Yuxuan Wang, Cheng Deng

Overview of the Remember Smarter visual-history and hyperbolic experience-memory branches — Remember Smarter combines fine-grained visual-history compression with coarse-grained hierarchical experience retrieval.

Abstract

Long-horizon robot policies need compact access to recent observations and reusable experience without continually expanding the vision-language-action context. Remember Smarter (RS) is a plug-and-play module with two complementary branches: a Mamba-based compressor for multi-view visual history and a hierarchical experience memory built in Poincaré space. The compressed history is exposed to action-facing states through residual cross-attention, while retrieved experience is converted asynchronously into geodesic prompt tokens. When integrated with π₀, RS improves total LIBERO-Plus success from 53.6% to 70.6% and delivers substantial gains on real-robot tasks that require memory retention and experience reuse.

Approach

Visual History Compressor

Bidirectional spatial Mamba encodes each multi-view frame, while causal temporal Mamba maintains an online recurrent history. Residual cross-attention lets action-facing hidden states read this memory without modifying the VLM visual-token stream.

Mamba-compressed visual-history pipeline

Hyperbolic Experience Space

Successful final-layer VLM states are embedded by a Poincaré VAE and organized hierarchically. Retrieved experiences are synthesized into geodesic prompt tokens without blocking action inference.

Poincare VAE experience space and HypHC hierarchy

Results

53.6 → 70.6 LIBERO-Plus total success (%)

20 → 80 Chenchen-hard success (%)

94.1 → 96.2 Standard LIBERO overall (%)

Component ablation across four standard LIBERO suites — The visual compressor drives the largest gain on LIBERO-Long; hierarchical experience memory further improves the full-system result.

Successful Remember Smarter rollout and baseline failure sequences on the Chenchen-hard task — Representative real-robot rollout: RS completes the compositional Chenchen-hard sequence, while the baseline fails at different subtasks.

BibTeX

@misc{zhou2026remembersmarter,
  title         = {Remember Smarter: Visual History Compressor and Hyperbolic Experience Space for Robotic Memory},
  author        = {Zhou, Dai and Yan, Jiexi and Li, Tong and Wang, Yuxuan and Deng, Cheng},
  year          = {2026},
  eprint        = {2608.15269},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2608.15269}
}