Towards Root Memories: Benchmarking and Enhancing Implicit Logical Memory Retrieval for Personalized LLMs

Personalized Large Language Models (LLMs) rely on memory systems to store user preferences and history. Current retrieval systems typically use semantic similarity (e.g., vector embeddings) to find relevant information. However, this approach often fails when a query is semantically distant from a critical piece of user history, even if that history is logically essential for a correct, personalized response. For example, a user might ask about buying a fish tank; a semantic retriever might find a memory about 'tropical fish' but miss a logically critical, semantically distant memory about the user's cat that makes an open-top tank dangerous.

To quantify this failure, the authors introduce IMLogic, the first benchmark specifically designed to evaluate implicit logical memory retrieval. The benchmark consists of 20 long-form conversations, over 15,000 memory entries, and 2,216 high-quality QA pairs. It explicitly models scenarios where semantic similarity and logical relevance diverge. The authors demonstrate that even state-of-the-art memory agents struggle with these tasks, often falling into 'preference traps' (pandering to the user's surface intent) or 'fabrication traps' (inventing details to justify a response).

RootMem is a plug-and-play framework designed to bridge this gap. It operates in two stages:

1. Distillation: It processes raw historical memories into structured 'root memory units.' Each unit contains execution rules (when to apply the logic) and personalized logical evidence (the facts supporting that logic).
2. Routing: During inference, an LLM-based router identifies which root memory units are relevant to the current query. These units are then combined with standard semantic retrieval results to provide a more robust, logic-aware context for the LLM.

Experiments show that RootMem significantly outperforms existing retrieval baselines, achieving a 27.23% relative improvement in accuracy on the IMLogic benchmark and consistently boosting the performance of existing memory agents.