Retrieved memory beats full history for LLM agents

Arxiv CS.CL (NLP/language models)

Two new works argue that providing LLM agents with a lean, retrieved context from a memory store yields better accuracy than feeding the full conversation history. Liuyin Wang et al. show a bi-temporal memory engine achieves higher accuracy with less context. Sonam Pankaj's work proposes utility-ranked memory to close the agent learning loop.

AgentsMemoryRetrieval

Arxiv CS.CL (NLP/language models)·Liuyin Wang

Jun 10, 4:00 AM