Researchers in Franklin College of Arts and Science have built a new memory architecture for AI agents that improves response accuracy on complex tasks up to 23 percent.  

The system, called Synapse (Synergistic Associative Processing and Semantic Encoding), addresses the current struggle for large-language models and agents to recall past conversations. This “contextual isolation” often results in weak and conflicting responses.  

For example, if someone asks an AI agent why he or she is feeling anxious, some models might offer tips and overlook a prior conversation about a recent interpersonal conflict. To combat this, UGA researchers have built their own memory architecture that reimagines memory for AI. Inspired by the human brain, Synapse prioritizes memories that are structurally important to a current conversation, even if some of the semantics are not identical or explicitly mentioned. This means that AI agents will be able to connect a weeks- or months-old interaction to a query in real time by focusing on reasoning rather than storage management. According to the researchers, this can improve the accuracy of responses to complex tasks by up to 23%. 

“The more (Synapse) learns, the better it gets,” said Professor Tianming Liu, Distinguished Research Professor of Computer Science. Liu’s research focuses on large language models and brain-inspired artificial intelligence. 

Liu and doctoral student Hanqi Jiang were two of several UGA authors attributed to the Synapse study, which connected disciplines at Franklin, including computer science and physics. Earlier this year, the research was accepted and presented at the Association for Computational Linguistics annual conference, which showcases premier research for language processing and AI. Doctoral students Junhao Chen, Yi Pan, Weihang You, Yifan Zhou, Ruidong Zhang, and physics professor Yohannes Abate also contributed to the research. 

Jiang said that Synapse will provide direct support to EZCollegeApp, a U.S. college application platform founded by Liu to help high school students track and organize multiple applications being completed at once.  

"An application season runs for months and looks exactly like a long-horizon memory benchmark, except the stakes are a teenager's future,” he said. 

Specifically, the AI agent must remember the student’s academic records across several weeks, and any advice it gives has to be grounded within those records, such as their transcripts, standardized test scores, or essay drafts.  

In addition to memory improvements, Synapse was built to help reduce AI hallucinations, or instances where an artificial intelligence system generates information or responses that are incorrect, misleading, or entirely fabricated. According to researchers, Synapse is designed to “fail safely.” In situations where it is uncertain, or if there is insufficient evidence, the system will refuse to answer.  

This will become increasingly important, Jiang said, especially in the medical field. AI researchers at UGA and other universities are working together to study how AI can be used to help clinicians interpret medical images. As models move into more real-world, high-stakes settings, Jiang and Liu said the trustworthiness of models will rely heavily on their ability to recognize what to consider from previous queries and when to safely refuse to answer if inputs have been unreliable or ambiguous.