LSG: A Lightweight Supervised Gate for Sub-Millisecond Agentic Memory Novelty Filtering

Abstract

This paper presents LSG (Lightweight Supervised Gate), a shallow calibrated ensemble architecture for write-time memory novelty filtering in long-horizon LLM agents. Contemporary agentic memory systems rely on cosine similarity thresholds for filtering incoming conversational turns — a heuristic that suffers from O(N) computational scaling, single-turn context blindness, and miscalibrated confidence estimates. LSG addresses these structural limitations by combining frozen 384-dimensional Sentence-BERT embeddings, multi-turn context history windowing (K=2), conversational metadata features, Platt probability calibration, and a 2nd-stage Logistic Regression Meta-Learner. Evaluated zero-shot on the PersonaChat benchmark (1,600 turns) against gold-standard human intent labels verified via LLM-as-a-Judge (κ = 0.84), LSG achieves competitive discriminative ranking (AUPRC = 0.569, ROC-AUC = 0.566) while executing in constant O(1) time at 0.15 ms on CPU — over 500× faster than Small Language Model alternatives such as Qwen2.5-0.5B and Llama-3.2-1B, and eliminating the memory bloat caused by degenerate pass-through heuristics. Keywords: agentic memory, memory write gating, LLM agents, Platt calibration, multi-turn context, Pareto efficiency, lightweight classifiers, RAG, conversational AI