Saved in:
Bibliographic Details
Main Authors: Lei, Jingdi, Zhang, Di, Li, Junxian, Wang, Weida, Fan, Kaixuan, Liu, Xiang, Liu, Qihan, Ma, Xiaoteng, Chen, Baian, Poria, Soujanya
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2605.12357
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914559755812864
author Lei, Jingdi
Zhang, Di
Li, Junxian
Wang, Weida
Fan, Kaixuan
Liu, Xiang
Liu, Qihan
Ma, Xiaoteng
Chen, Baian
Poria, Soujanya
author_facet Lei, Jingdi
Zhang, Di
Li, Junxian
Wang, Weida
Fan, Kaixuan
Liu, Xiang
Liu, Qihan
Ma, Xiaoteng
Chen, Baian
Poria, Soujanya
contents Large language models increasingly need to accumulate and reuse historical information in long-term assistants and agent systems. Simply expanding the context window is costly and often fails to ensure effective context utilization. We propose $δ$-mem, a lightweight memory mechanism that augments a frozen full-attention backbone with a compact online state of associative memory. $δ$-mem compresses past information into a fixed-size state matrix updated by delta-rule learning, and uses its readout to generate low-rank corrections to the backbone's attention computation during generation. With only an $8\times8$ online memory state, $δ$-mem improves the average score to $1.10\times$ that of the frozen backbone and $1.15\times$ that of the strongest non-$δ$-mem memory baseline. It achieves larger gains on memory-heavy benchmarks, reaching $1.31\times$ on MemoryAgentBench and $1.20\times$ on LoCoMo, while largely preserving general capabilities. These results show that effective memory can be realized through a compact online state directly coupled with attention computation, without full fine-tuning, backbone replacement, or explicit context extension.
format Preprint
id arxiv_https___arxiv_org_abs_2605_12357
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle $δ$-mem: Efficient Online Memory for Large Language Models
Lei, Jingdi
Zhang, Di
Li, Junxian
Wang, Weida
Fan, Kaixuan
Liu, Xiang
Liu, Qihan
Ma, Xiaoteng
Chen, Baian
Poria, Soujanya
Artificial Intelligence
Large language models increasingly need to accumulate and reuse historical information in long-term assistants and agent systems. Simply expanding the context window is costly and often fails to ensure effective context utilization. We propose $δ$-mem, a lightweight memory mechanism that augments a frozen full-attention backbone with a compact online state of associative memory. $δ$-mem compresses past information into a fixed-size state matrix updated by delta-rule learning, and uses its readout to generate low-rank corrections to the backbone's attention computation during generation. With only an $8\times8$ online memory state, $δ$-mem improves the average score to $1.10\times$ that of the frozen backbone and $1.15\times$ that of the strongest non-$δ$-mem memory baseline. It achieves larger gains on memory-heavy benchmarks, reaching $1.31\times$ on MemoryAgentBench and $1.20\times$ on LoCoMo, while largely preserving general capabilities. These results show that effective memory can be realized through a compact online state directly coupled with attention computation, without full fine-tuning, backbone replacement, or explicit context extension.
title $δ$-mem: Efficient Online Memory for Large Language Models
topic Artificial Intelligence
url https://arxiv.org/abs/2605.12357