EPGM: Execution Provenance for Budgeted Agent Memory Retrieval (arxiv.org)

A new framework called EPGM (Execution Provenance for Budgeted Agent Memory Retrieval) has been introduced to enhance the efficiency of memory retrieval for language agents. This approach addresses significant limitations in current systems that utilize fixed token windows for retrieving contextual evidence, which can lead to the scattering of relevant information and reduced performance. By formulating memory retrieval as a budgeted evidence completion task, EPGM constructs source-aligned provenance units from tool arguments and outputs, optimizing the retrieval process to ensure that all relevant evidence fits within a specified token budget. The significance of EPGM lies in its potential to improve the accuracy of memory retrieval in multi-agent systems, particularly when handling complex tasks that require evidence spanning multiple events. The framework employs a zero-initialized residual Graph Convolutional Network (R-GCN) that refines retrieval scores based on typed provenance edges, showing a notable improvement in performance metrics. In evaluations, EPGM achieved an increase of 19.07 points in Full Support@2048 compared to traditional methods, demonstrating its effectiveness in managing evidence accumulation without overfragmentation. This advancement highlights the importance of leveraging graph structures in retrieval tasks, promising better support for agents as they navigate intricate execution histories in AI and ML applications.