People continually encounter opinions through news media and social networks. They decide what to accept, remember, reject, or share. Over time, these choices can reshape individual worldviews and generate collective patterns such as consensus, polarization, and filter bubbles. We develop an agent-based model that links these micro-level information choices to macro-level opinion dynamics. Rather than representing opinions as points in a low-dimensional numerical space, our approach uses natural-language opinion statements and LLM agents that can understand and act on them. Each agent has a bounded memory of statements representing its worldview. Agents encounter statements from a shared feed and network neighbors, decide whether to integrate or reject them, replace existing statements when memory is full, and share remembered statements with others. Unlike persona-based simulations, agents are not assigned demographic profiles or ideological labels. Their worldviews develop through repeated exposure, integration, replacement, and sharing. Preliminary results produce collective patterns, including convergence around common statements, concentration of attention on a narrow subset of the statement pool, and separation into rival worldview clusters. Early comparisons suggest that memory capacity may shape whether collective attention narrows around a few statements or preserves diversity for competing clusters. These findings connect individual information choices to emergent population level worldviews.