LAMB Framework for Latent Multi-Hypothesis Beliefs in Long-Horizon Agents

LAMB uses a Common-Delta factorization to maintain a shared interaction state alongside weighted delta components for competing hypotheses. This recurrent framework avoids textual memory compression bottlenecks by updating latent beliefs through action-outcome prediction without generating natural language summaries.