ElasticMem: Latent Memory as a Learnable Resource for LLM Agents

Tao Feng 1 , Chongrui Ye 1 , Tianyang Luo 1 , Jingjun Xu 1 , Xueqiang Xu 1 , Haozhen Zhang 2 Ge Liu 1 , Jiaxuan You 1 1 University of Illinois Urbana-Champaign 2 Nanyang Technological University · 2026

Read ElasticMem: Latent Memory as a Learnable Resource for LLM Agents by Tao Feng 1 , Chongrui Ye 1 , Tianyang Luo 1 , Jingjun Xu 1 , Xueqiang Xu 1 , Haozhen Zhang 2 Ge Liu 1 , Jiaxuan You 1 1 University of Illinois Urbana-Champaign 2 Nanyang Technological University (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-30690.txt.

Description

Feng and colleagues present ElasticMem, treating memory as a learnable latent resource for language-model agents. Old approaches fix memory allocation: text methods stuff memories into context, while latent methods still use rigid retrieval. ElasticMem builds a latent memory bank and retrieves from the reasoner's hidden state. It assigns each memory a budget through a learned policy, and injects selected states as soft tokens. Optimized through group-relative policy optimization, it improves question-answering accuracy and control success on two backbones while using the fewest tokens on ALFWorld.

Catalog record

CREATORTao Feng 1 , Chongrui Ye 1 , Tianyang Luo 1 , Jingjun Xu 1 , Xueqiang Xu 1 , Haozhen Zhang 2 Ge Liu 1 , Jiaxuan You 1 1 University of Illinois Urbana-Champaign 2 Nanyang Technological University
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE137112
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.30690

Browse the full library · About the Tower of Babel