Adaptive Memory Admission Control for LLM Agents
arXiv:2603.04549v1 Announce Type: new Abstract: LLM-based agents increasingly rely on long-term memory to support multi-session reasoning and interaction, yet current systems provide little control over what information is retained. In practice, agents either accumulate large volumes of conversational content, including hallucinated or obsolete facts, or depend on opaque, fully LLM-driven memory policies that are costly and difficult to audit. As a result, memory admission remains a poorly specified and weakly controlled component in agent architectures. To address this gap, we propose Adaptive Memory Admission Control (A-MAC), a framework that treats memory admission as a structured decision problem. A-MAC decomposes memory value into five complementary and interpretable factors: future utility, factual confidence, semantic novelty, temporal recency, and content type prior. The framework combines lightweight rule-based feature extraction with a single LLM-assisted utility assessment,
arXiv:2603.04549v1 Announce Type: new Abstract: LLM-based agents increasingly rely on long-term memory to support multi-session reasoning and interaction, yet current systems provide little control over what information is retained. In practice, agents either accumulate large volumes of conversational content, including hallucinated or obsolete facts, or depend on opaque, fully LLM-driven memory policies that are costly and difficult to audit. As a result, memory admission remains a poorly specified and weakly controlled component in agent architectures. To address this gap, we propose Adaptive Memory Admission Control (A-MAC), a framework that treats memory admission as a structured decision problem. A-MAC decomposes memory value into five complementary and interpretable factors: future utility, factual confidence, semantic novelty, temporal recency, and content type prior. The framework combines lightweight rule-based feature extraction with a single LLM-assisted utility assessment, and learns domain-adaptive admission policies through cross-validated optimization. This design enables transparent and efficient control over long-term memory. Experiments on the LoCoMo benchmark show that A-MAC achieves a superior precision-recall tradeoff, improving F1 to 0.583 while reducing latency by 31% compared to state-of-the-art LLM-native memory systems. Ablation results identify content type prior as the most influential factor for reliable memory admission. These findings demonstrate that explicit and interpretable admission control is a critical design principle for scalable and reliable memory in LLM-based agents.
Executive Summary
The article proposes Adaptive Memory Admission Control (A-MAC), a framework for controlling long-term memory in LLM-based agents. A-MAC decomposes memory value into five factors and combines rule-based feature extraction with LLM-assisted utility assessment to learn domain-adaptive admission policies. Experiments show that A-MAC achieves a superior precision-recall tradeoff and reduces latency compared to state-of-the-art systems. The framework enables transparent and efficient control over long-term memory, demonstrating the importance of explicit and interpretable admission control for scalable and reliable memory in LLM-based agents.
Key Points
- ▸ A-MAC framework for controlling long-term memory in LLM-based agents
- ▸ Decomposition of memory value into five complementary factors
- ▸ Combination of rule-based feature extraction and LLM-assisted utility assessment
Merits
Improved Precision-Recall Tradeoff
A-MAC achieves a superior precision-recall tradeoff, improving F1 to 0.583, compared to state-of-the-art LLM-native memory systems.
Reduced Latency
A-MAC reduces latency by 31% compared to state-of-the-art LLM-native memory systems.
Demerits
Dependence on LLM-Assisted Utility Assessment
A-MAC relies on a single LLM-assisted utility assessment, which may be costly and difficult to audit.
Expert Commentary
The proposed A-MAC framework represents a significant step forward in addressing the challenges of memory admission control in LLM-based agents. By decomposing memory value into interpretable factors and combining rule-based feature extraction with LLM-assisted utility assessment, A-MAC provides a transparent and efficient approach to controlling long-term memory. The experimental results demonstrate the effectiveness of A-MAC in achieving a superior precision-recall tradeoff and reducing latency. However, the reliance on LLM-assisted utility assessment may be a limitation, and future research should explore alternative approaches to mitigate this dependence.
Recommendations
- ✓ Future research should explore the application of A-MAC to various LLM-based agents and domains.
- ✓ The development of A-MAC should be accompanied by the creation of standards and guidelines for transparent and explainable AI system design.