ROAM is a framework for managing atomic memories in language model agents. It classifies atom pairs as independent, equivalent, directionally subsuming, or conflicting. This classification informs the organization of observations into active Primary and supporting Evidence roles. The system then fuses complementary details and temporal changes into compact views, prioritizing Primary views for answering. This approach reduces redundant or outdated atoms from competing independently. Across models and evaluation settings, ROAM improves answer accuracy by up to 29.8 percentage points. Mechanism analysis shows 15.6-point higher answer-critical source recall and an 11.5-point lower confounder-token share. ROAM remains robust across manager scales.
Source: https://arxiv.org/abs/2609.09778