Towards Root Memories: Benchmarking and Enhancing Implicit Logical Memory Retrieval for Personalized LLMs
2026-06-22 • Computation and Language
Computation and Language
AI summaryⓘ
The authors point out that current memory systems in personalized large language models mostly find past information based on word similarity, which can miss important logical details. To fix this, they created IMLogic, a new test set that focuses on finding logical memories in long conversations. They also developed RootMem, a method that organizes past user information into clear logical units called root memories and uses a smart system to pick the right ones. Their tests show RootMem works better than older methods at retrieving important memories and improving model accuracy.
Large Language Modelsmemory retrievalsemantic similaritylogical memorybenchmark datasetlong-dialogue scenariosstructured representationdecision logicmemory agentsrouting
Authors
Hongxun Ding, Xiang Yu, Chengbing Wang, Jianfei Xiao, Keqin Bao, Wenjie Wang, Xiangnan He
Abstract
Memory systems are essential for personalized Large Language Models (LLMs). However, existing retrieval methods in these systems primarily rely on semantic similarity, potentially missing logically critical memories with limited semantic overlap. Current benchmarks remain inadequate for evaluating this problem. To address this gap, we construct IMLogic, the first high-quality benchmark targeting implicit logical memory retrieval in long-dialogue scenarios. Motivated by this challenge, we introduce root memory, a structured, decision-preserving representation that distills reusable personalized logic from long-term user histories. We then propose RootMem, a plug-and-play framework that first distills raw histories into structured root memories and then uses an LLM-based router to activate logically relevant ones, complementing semantic retrieval with personalized decision logic. Extensive experiments demonstrate that RootMem significantly outperforms the strongest retrieval baselines and consistently boosts the accuracy of existing memory agents. Our benchmark and codes will be available at https://anonymous.4open.science/r/IMLogic-DBB3.