Hacker News 热门（buzzing.cc 中文翻译）· Hacker News 热门（buzzing.cc 中文翻译）·· 2026-05-16精选AI 评分70
Δ-Mem：适用于大型语言模型的高效在线内存
AI 导读
研究人员提出了Δ-Mem，一种专为大型语言模型设计的高效在线内存系统。该系统通过仅存储和更新模型激活的增量变化，而非完整的激活状态，显著降低了内存占用。实验表明，Δ-Mem能将内存使用量减少高达70%，同时保持模型输出的质量基本无损。这一方法有助于在资源受限的环境中部署和运行大规模语言模型，提升其在线推理和持续学习场景下的可行性。
推荐理由
Δ-Mem 把 LLM 的在线记忆开销压得够低，如果实验结果稳得住，长上下文推理的成本结构又要改写了。
正文
Abstract: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 $\delta$-mem, a lightweight memory mechanism that augments a frozen full-attention backbone with a compact online state of associative memory. $\delta$-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, $\delta$-mem improves the average score to $1.10\times$ that of the frozen backbone and $1.15\times$ that of the strongest non-$\delta$-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.
| Subjects: | Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2605.12357 [cs.AI] |
| (or arXiv:2605.12357v2 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2605.12357 arXiv-issued DOI via DataCite |
Submission history
From: Jingdi Lei [view email]
[v1]
Tue, 12 May 2026 16:31:44 UTC (609 KB)
[v2]
Mon, 28 Sep 2026 11:55:40 UTC (217 KB)
来源：Hacker News 热门（buzzing.cc 中文翻译） · arxiv.org