[Paper Review] Scene-Aware Memory Discrimination: Deciding Which Personal Knowledge Stays

Yijie Zhong, Mengying Guo+4 more|arXiv (Cornell University)|Feb 12, 2026

Personal Information Management and User Behavior0 citations

TL;DR

SAMD uses a Gating Unit and Cluster Prompting Module to perform scene-aware memory discrimination, improving memory quality while reducing computation, with strong direct and indirect evaluations.

ABSTRACT

Intelligent devices have become deeply integrated into everyday life, generating vast amounts of user interactions that form valuable personal knowledge. Efficient organization of this knowledge in user memory is essential for enabling personalized applications. However, current research on memory writing, management, and reading using large language models (LLMs) faces challenges in filtering irrelevant information and in dealing with rising computational costs. Inspired by the concept of selective attention in the human brain, we introduce a memory discrimination task. To address large-scale interactions and diverse memory standards in this task, we propose a Scene-Aware Memory Discrimination method (SAMD), which comprises two key components: the Gating Unit Module (GUM) and the Cluster Prompting Module (CPM). GUM enhances processing efficiency by filtering out non-memorable interactions and focusing on the salient content most relevant to application demands. CPM establishes adaptive memory standards, guiding LLMs to discern what information should be remembered or discarded. It also analyzes the relationship between user intents and memory contexts to build effective clustering prompts. Comprehensive direct and indirect evaluations demonstrate the effectiveness and generalization of our approach. We independently assess the performance of memory discrimination, showing that SAMD successfully recalls the majority of memorable data and remains robust in dynamic scenarios. Furthermore, when integrated into personalized applications, SAMD significantly enhances both the efficiency and quality of memory construction, leading to better organization of personal knowledge.

Motivation & Objective

• Motivate memory discrimination to build concise, high-quality personal memories for personalized applications.
• Propose a scene-aware memory discrimination framework (SAMD) that operates with a frozen LLM.
• Design GUM to filter non-memorable data using memory-scene saliency and CPM to define memory standards via intent-scene clustering.
• Demonstrate SAMD’s effectiveness and robustness through direct and indirect evaluations across diverse datasets.

Proposed method

• Introduce memory discrimination as a binary decision: whether each sentence in context should be stored as memory.
• Gating Unit Module (GUM) filters non-memorable data using a memory scene-based identifier built from salient words and multi-view role-playing.
• Cluster Prompting Module (CPM) defines what to remember by clustering user intents with an intent-scene affinity matrix and building discrimination rules per cluster.
• Use a frozen LLM to apply discrimination rules; update accessibility to new memory scenes and intents with minimal cost.
• Directly evaluate memory discrimination accuracy, recall, and efficiency; indirectly evaluate impact on memory construction for memory-enabled agents.

Experimental results

Research questions

• RQ1 Can SAMD accurately discriminate memorable versus non-memorable data at scale?
• RQ2 Does scene-aware discrimination improve memory quality and efficiency across varied memory scenes and intents?
• RQ3 How well does SAMD adapt to new memory scenes or intents without fine-tuning the LLM?
• RQ4 What is the impact of SAMD on downstream memory-assisted applications in terms of speed and quality?

Key findings

• SAMD with intent clustering achieves about 3x faster discrimination and over 20% higher accuracy than baseline methods on key datasets.
• SAMD retains more than 90% of memorable data, improving overall performance by over 20%.
• GUM effectively filters non-memorable data early, reducing unnecessary LLM usage and computational cost.
• CPM’s intent-scene clustering enables discrimination rules that cover diverse expressions while minimizing memory gaps across scenes.
• SAMD remains robust when user intents are incomplete, incorrect, or unavailable, and adapts to evolving application demands.