[Submitted connected 31 Jul 2026]
Authors:Yilin Xiao, Zhehan Zhu, Yujing Zhang, Jin Chen, Zijin Hong, Luyao Zhuang, Qinggang Zhang, Shengyuan Chen, Xiaocao Ouyang, Lingfei Ren, Xiao Huang
View PDF HTML (experimental)
Abstract:LLM agents request representation to enactment consistently complete agelong interactions, yet galore systems usage further LLM calls to run that memory. Generating intermediate records and mediating their retrieval adds recurring token and clip costs, while omitted aliases merged specifications tin obscure the original evidence. We inquire whether system representation entree requires procreation astatine all. Zero-Mem introduces \emph{zero-token representation operations}: nary measurement extracurricular last mobility answering invokes an LLM aliases consumes LLM input aliases output tokens; encoder computation is accounted for separately. Zero-Mem preserves original relationship traces arsenic its root of record. It organizes the traces successful 2 complementary ways. An entity--context chart exposes connections crossed interactions, while a temporal level preserves conversational locality and convention state. For each query, Zero-Mem weighs the 2 views, retrieves from both, and follows their building to retrieve supporting relations aliases surrounding context. Deterministic calibration first discards conflicting grounds and past keeps the reader's reply grounded successful the retrieved traces. Only the final-QA scholar invokes an LLM. Across long-memory and long-context question-answering benchmarks, Zero-Mem achieves competitory capacity while eliminating LLM calls and LLM-token depletion from representation operations. With the aforesaid final-QA scholar and discourse budget, it reduces memory-operation clip costs by 57.6\% comparative to the fastest compared baseline. Ablations support the publication of the 2 views and their query-dependent coordination. Overall, the results show that system supplier representation request not make an intermediate practice of the past. After adjacent review, the codification and implementation specifications will beryllium disposable astatine \textcolor{blue}{this https URL}.Submission history
From: Yilin Xiao [view email]
[v1] Fri, 31 Jul 2026 13:01:06 UTC (414 KB)
English (US) ·
Indonesian (ID) ·