Multi-Layered Memory Architectures for LLM Agents: An Experimental Evaluation of Long-Term Context Retention

TL;DR

This paper introduces a Multi-Layer Memory Framework for long-horizon dialogue systems, improving long-term context retention and reasoning stability through layered memory management and adaptive retrieval.

Contribution

The paper proposes a novel layered memory architecture with gating and regularization, effectively reducing semantic drift and enhancing long-term memory in dialogue agents.

Findings

Achieved 46.85 Success Rate and 0.618 overall F1 on benchmark datasets.

Reduced false memory rate to 5.1% and context usage to 58.40%.

Demonstrated improved long-term retention and reasoning stability.

Abstract

Long-horizon dialogue systems suffer from semanticdrift and unstable memory retention across extended sessions. This paper presents a Multi-Layer Memory Framework that decomposes dialogue history into working, episodic, and semantic layers with adaptive retrieval gating and retention regularization. The architecture controls cross-session drift while maintaining bounded context growth and computational efficiency. Experiments on LOCOMO, LOCCO, and LoCoMo show improved performance, achieving 46.85 Success Rate, 0.618 overall F1 with 0.594 multi-hop F1, and 56.90% six-period retention while reducing false memory rate to 5.1% and context usage to 58.40%. Results confirm enhanced long-term retention and reasoning stability under constrained context budgets.