← News

Agents & MCPSources: arXiv · 20 Nov 2026

FlashX stores robot experience in NAND flash to speed unfamiliar tasks

A research team proposes a flash-memory architecture that lets embodied agents reuse past successful actions, reporting accuracy and speed gains over SSD-based retrieval.

In plain words

Robots that act from vision and language often fail in situations they were not trained for. FlashX is a proposed memory design that keeps a bank of past successful robot actions on flash storage and searches it quickly, so the robot can start from a known-good action instead of computing one from scratch. The authors say this improved both task success and speed in their tests.

What was announced

The arXiv abstract describes embodied agents that generate actions through Vision-Language-Action models and World Action Models using Diffusion Transformer-based iterative denoising, and notes these models often struggle in out-of-distribution environments. The authors observe that retrieving successful experiences to warm-start denoising can help agents adapt to unfamiliar situations while lowering computational overhead, unlike conventional acceleration techniques that trade accuracy for computation. Scaling such experience banks, however, creates storage and search overhead that strains the capacity, bandwidth, and power limits of edge systems.

FlashX addresses this with a CMOS Directly Bonded to Array NAND flash architecture. It uses the high internal bandwidth of CBA flash for parallel pSLC key search through interleaved Compute-in-Flash engines, while dense TLC storage holds latents, and proprioception-aware clustering reduces search overhead and allows new experiences to be added without global reclustering. In evaluations on VLA and WAM workloads, the authors report up to 12.24 percentage points and 7.43 percentage points higher task success under dense and compute-efficient configurations, 2.57 times higher search throughput than the strongest prior NAND search architecture, a further 9.7 times search reduction from clustering, and 1.29 to 3.95 times end-to-end speedup with 1.12 to 2.34 times better energy efficiency than SSD-based retrieval.

What it does not show

  • Evidence is limited to the paper's abstract; experimental setup, baselines, and peer review are not described, and the reported gains are the authors' own claims not independently confirmed.

Sources

Source-support review completed 21 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 21 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.