Reproduce “Beyond RAG for Agent Memory: Retrieval by Decoupling and Aggregation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhanghao Hu, Qinglin Zhu, Runcong Zhao, Di Liang, Hanqi Yan, Yulan He, Lin Gui
xMemory builds long-term dialogue memory by extracting small facts from conversation segments, linking them to sources, and grouping related facts before retrieval. It then selects complementary groups and adds source text when it reduces uncertainty. On LoCoMo with Qwen3-8B, it scores 34.48/43.98 aggregate BLEU-1/F1 versus Nemori’s 31.44/40.88, but results are benchmark-specific and construction adds cost.
Key insights
- The "decoupling before aggregation" principle effectively addresses redundancy and detail loss in agent memory by explicitly isolating and then structuring fine-grained evidence units.
- Dynamically maintaining the memory structure through split and merge operations is crucial for adapting to evolving agent interactions and preserving memory integrity over time.
- Adaptive two-stage retrieval, driven by LLM uncertainty reduction, delivers denser and more concentrated evidence, leading to improved answer quality with optimized token efficiency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.