Membox: Weaving Topic Continuity into Long-Range Memory for LLM Agents

A new memory architecture called Membox organizes long-running AI dialogues by topic continuity rather than isolated turns, yielding substantial retrieval gains over existing systems, according to a paper posted on arXiv [1]. The work, authored by Dehao Tao and colleagues, introduces a hierarchical design that treats topic continuity as an explicit organizational layer for agent memory [1]. The system’s Topic Loom incrementally groups dialogue streams into boxes whose internal turns follow the same local topic, while a second component, the Trace Weaver, links extracted events across boxes into macro-topic traces that recover recurring activities, goals, and factual developments across distant sessions [1]. The paper argues that many current LLM agent memory systems first decompose histories into isolated turns or fixed-size chunks, then rely on enrichment, consolidation, or retrieval mechanisms still tied to semantic proximity or fragment-level records [1]. This approach, the authors write, “weakens temporal and causal organization and biases memory access toward semantic proximity rather than task- or topic-level continuity” [1]. On the LoCoMo dataset, Topic-Loom-only retrieval improved over the best Mem0/A-MEM retrieval-depth setting by 13.00 F1 points, reaching 53.95 F1 compared with 40.95 [1]. When trace-expanded retrieval was added, the score rose to 55.28 [1]. With GPT-4o, trace-expanded retrieval reached 59.71 F1 [1]. Additional experiments on the DialSim dataset showed the same gain from adding cross-box traces in multi-party dialogue [1]. The paper was submitted to arXiv on January 7, 2026, and revised most recently on June 24, 2026 [1]. arXiv, an open-access repository of electronic preprints and postprints, was founded in 1991 and now receives about 24,000 submissions per month as of November 2024 [6]. The platform hosts papers across mathematics, physics, computer science, and related fields, and its content is moderated but not peer-reviewed [6]. The Membox paper appears with experimental tools provided through arXivLabs, a framework launched in 2020 that allows community collaborators to develop and share features directly on the arXiv site [4]. arXivLabs projects include bibliographic explorers, code-and-data linkers, and recommender systems, all governed by guidelines that emphasize openness, community, excellence, and user data privacy [4][5]. The framework is currently on hiatus for new proposals while the arXiv development team focuses on modernizing and migrating systems to the cloud [3].