Source-linked AI summary
Adaptive Memory Admission Control for LLM Agents
Guilin Zhang, Wei Jiang, Xiejiashan Wang, Aisha Behr, Kai Zhao, Jeffrey Friedman, Xu Chu, Amine Anoun
TL;DR
Long-term memory is essential for multi-session LLM agents, but existing admission policies provide limited control over hallucinated, obsolete, or excessive content. A-MAC frames admission as a structured decision using five interpretable factors and a hybrid rule-based/LLM design, achieving a stronger precision–recall tradeoff with lower latency on LoCoMo. Its results support explicit, adaptive admission control as a practical approach to long-term memory management.
Problem
Memory admission in LLM agents is poorly specified, with existing systems either retaining excessive or unreliable content or relying on costly, opaque LLM-driven policies.
Method
A-MAC scores candidate memories using utility, confidence, novelty, recency, and type prior, combining rule-based features with one LLM-assisted utility assessment and cross-validated policy optimization.
Results
A-MAC reaches F1=0.583 on LoCoMo and runs 31% faster than prior methods, with Type Prior identified as the dominant feature.
Takeaways & Limitations
Explicit, interpretable admission control provides adaptive and efficient control over which information is promoted to long-term memory.
Abstract
from arXiv · showhide
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.
1 INTRODUCTION
Memory admission is an under-specified control problem because existing heuristic and LLM-driven approaches trade off reliability, interpretability, and efficiency. A-MAC addresses this by evaluating candidate memories through five interpretable dimensions and combining rule-based features with minimal LLM assistance.
- A-MAC Framework: A-MAC treats memory admission as a structured decision problem before information enters long-term storage.This makes admission an explicit control mechanism for balancing memory coverage, reliability, and efficiency.
- A-MAC Framework: A-MAC evaluates memory value using future utility, factual confidence, semantic novelty, temporal recency, and content type prior.These dimensions respectively address future relevance, evidential support, redundancy, temporal decay, and domain-specific persistence preferences.
- Motivation: Memory admission remains under-specified, while heuristic and fully LLM-driven approaches have important limitations.Heuristic methods struggle with subtle decisions and hallucination prevention, whereas LLM-native methods impose computational overhead and limited auditability.
- Method: The framework combines rule-based feature computation with minimal LLM inference to balance interpretability, accuracy, and computational efficiency.Admission policies are learned through cross-validated optimization to adapt across conversational domains without manual tuning.
- Results: A-MAC achieves a superior precision–recall tradeoff on LoCoMo while reducing latency relative to state-of-the-art LLM-native memory systems.Ablation studies examine the role of individual admission factors.
2 RELATED WORK
Prior memory systems either use efficient but rigid heuristics or flexible but costly and opaque LLM-native policies. A-MAC introduces a data-driven admission layer that combines interpretable criteria, selective LLM use, and learned weighting.
- Heuristic Memory Management: Hierarchical and heuristic systems emphasize efficient storage but rely on static or manually tuned admission rules.MemGPT lacks explicit factual verification, while MemoryBank uses fixed weights that cannot adapt across domains.
- LLM-Native Memory Management: LLM-native systems improve recall but require substantial computation and provide limited interpretability.A-mem uses multiple LLM invocations for structured attributes and similarity-based matching.
- Retrieval-Based Methods: Retrieval-augmented methods improve grounding and reasoning but primarily target retrieval quality rather than memory-growth control.Their focus differs from admission policies that decide what enters long-term storage.
- A-MAC Positioning: A-MAC bridges these approaches with selective LLM usage, rule-based reliability checks, and learned weighting at memory admission.Its design aims to provide efficient, interpretable, and hallucination-aware control over information entering long-term storage.
3 METHODOLOGY
A-MAC treats memory admission as an explicit decision process: candidate memories are normalized, scored across five interpretable signals, and processed by a learned threshold policy. The framework combines lightweight rules with limited LLM assistance and updates memory by admitting, rejecting, or merging candidates.
- Memory admission as a decision problem: A-MAC extracts semantically atomic candidate memories from conversation history and decides whether to admit, update, or reject each one.Candidates are normalized by segmenting turns, resolving temporal expressions and coreferences, and filtering low-value content.
- Admission scoring policy: S(m) combines the five feature values with nonnegative weights summing to one, and candidates are admitted when S(m) ≥ θ.The policy uses a learned scalar score and threshold to control admission decisions.
- Interpretable memory value signals: Each candidate receives five complementary signals: utility, confidence, novelty, recency, and content type prior.Together, these signals represent future usefulness, evidential support, redundancy, temporal decay, and persistence preferences.
- Policy learning and memory updates: Five-fold cross-validation learns the feature weights and threshold by maximizing F1 over nonnegative weight configurations and candidate thresholds.The selected policy computes all signals in parallel and merges a higher-scoring candidate when it conflicts with an existing memory.
4 RESULTS
On LoCoMo, A-MAC combines strong admission quality with lower latency, while ablations and cross-domain tests identify Type Prior as especially influential and show transfer without retuning.
- Main results: A-MAC achieves F1=0.583 on the LoCoMo test set, outperforming A-mem at 0.541 by 7.8% relative improvement.It also reaches precision 0.417 and recall 0.972.
- Latency analysis: 31% lower latency than A-mem, 2644ms versus 3831ms per candidate, demonstrates a substantial efficiency gain.A-MAC uses one LLM call while its four rule-based features execute in under 65ms combined.
- Ablation study: Removing Type Prior causes the largest ablation loss, with ∆F1=−0.107 and F1 falling from 0.583 to 0.476.The remaining features contribute smaller incremental losses ranging from −0.013 to −0.028.
- Threshold sensitivity: F1 peaks at θ∗=0.55 with validation F1 0.571, generalizing to 0.583 on the held-out test set.The F1 plateau from θ=0.50 to θ=0.60 suggests robustness to threshold selection.
- Cross-domain generalization: A-MAC achieves F1=0.482 on Personal and F1=0.338 on Professional conversations while retaining the same learned weights across domains.Personal conversations perform better because explicit preference statements align with Type Prior’s strengths.
5 CONCLUSION
A-MAC provides interpretable, efficient memory admission by combining five scored dimensions with learned weighting. Its results identify Type Prior as dominant and support inspection of admission decisions for debugging, auditing, and policy improvement.
- Conclusion: A-MAC scores candidate memories across five interpretable dimensions: Utility U, Confidence C, Novelty N, Recency R, and Type Prior T.Rule-based features C, N, R, and T are combined with LLM-based Utility scoring.
- Conclusion: A-MAC achieves state-of-the-art F1 (0.583) while running 31% faster than prior methods.Its linear weighted scoring model also exposes individual feature scores and weight contributions.
- Conclusion: Type Prior is the dominant feature, indicating that content category provides the strongest signal for memory admission decisions.The conclusion connects this result to interpretable inspection of why memories are admitted or rejected.