New Memory-Based Frameworks Enhance LLM Agent Learning and Reasoning

New Memory-Based Frameworks Enhance LLM Agent Learning and Reasoning

Researchers have introduced multiple memory-augmented frameworks aimed at improving how large language model (LLM) agents learn and reason without requiring parameter updates.

According to arxiv.org, one framework leverages both episodic and semantic memory to help LLM agents learn classification functions from labeled examples. The approach uses episodic memory to store instance-level critiques capturing specific past experiences, while semantic memory distills these into reusable task-level guidance. The best performing strategy yielded an average improvement of 8.1 percentage points over zero-shot baselines and 4.6 percentage points over RAG-based baselines. The framework also reduced thinking tokens by an average of 31.95% across datasets by using pre-computed critiques.

Separately, arxiv.org reports on E-mem, a multi-agent framework accepted at ICML 2026 that shifts from memory preprocessing to episodic context reconstruction. E-mem employs a hierarchical architecture where assistant agents maintain uncompressed memory contexts while a master agent orchestrates planning. On the LoCoMo benchmark, E-mem achieved over 54% F1 score, surpassing the state-of-the-art GAM by 7.75% while reducing token cost by over 70%.

A third system, HyMem, addresses the efficiency-effectiveness trade-off through hybrid memory architecture with dynamic retrieval scheduling, according to arxiv.org. HyMem uses dual-granular storage and a two-tier retrieval system, achieving strong performance on LOCOMO and LongMemEval benchmarks while reducing computational cost by 92.6%.