Researchers Advance LLM Agent Memory Systems with Episodic and Semantic Approaches

Researchers Advance LLM Agent Memory Systems with Episodic and Semantic Approaches

Several research teams have published new approaches to improving how large language model (LLM) agents learn and retain information, according to papers recently published on arxiv.org.

According to one arxiv.org paper, researchers developed a memory-augmented framework that enables LLM agents to learn classification tasks from labeled examples without parameter updates. The approach uses episodic memory to store instance-level critiques and semantic memory to distill reusable guidance. The best performing strategy “yields an average improvement of 8.1 percentage points over the zero shot baseline,” according to the paper, and reduces thinking tokens by an average of 31.95% across datasets.

Separately, researchers introduced E-mem, a multi-agent framework accepted by ICML 2026, according to arxiv.org. The system employs “a heterogeneous hierarchical architecture where multiple assistant agents maintain uncompressed memory contexts, while a central master agent orchestrates global planning.” E-mem achieved “over 54% F1, surpassing the state-of-the-art GAM by 7.75%, while reducing token cost by over 70%” on the LoCoMo benchmark, according to the paper.

Another arxiv.org paper described HyMem, a hybrid memory architecture with dynamic retrieval scheduling. The system uses “a dual-granular storage scheme paired with a dynamic two-tier retrieval system” and reportedly outperformed full-context approaches while reducing computational cost by 92.6% on LOCOMO and LongMemEval benchmarks.

Additionally, researchers released ViLegalNLI, the first large-scale Vietnamese Natural Language Inference dataset for legal domains, containing 42,012 premise-hypothesis pairs, according to arxiv.org.