Enhancing memory retrieval in generative agents through LLM-trained cross attention networks.

Area of Science:

• Artificial Intelligence
• Cognitive Science

Background:

• Large language models (LLMs) are advancing Artificial General Intelligence (AGI).
• Generative agents require unique memory retrieval systems to maintain individual characteristics.
• High costs of individual LLM training necessitate efficient memory management.

Purpose of the Study:

• To develop a novel memory retrieval system for generative agents.
• To enhance agent adaptability and behavioral consistency in dynamic environments.
• To explore the utility of LLMs in training dedicated agent memory retrieval networks.

Main Methods:

• Developed a text-based simulation of a generative agent community.
• Introduced an Auxiliary Cross Attention Network (ACAN) for memory retrieval.
• Utilized LLM assistance to compare and optimize memory retrieval through a novel loss function.

Main Results:

• Empirical evaluations show substantially enhanced memory retrieval quality.
• The approach significantly increases agent adaptability in fluctuating environments.
• Improved behavioral consistency of agents was observed.

Conclusions:

• The study presents a novel methodology for memory retrieval in generative agents.
• Findings extend the application of LLMs in AI agent memory management.
• This work pioneers the use of LLMs for training dedicated agent memory retrieval networks.