Forgetting helps AI agents agree on shared meanings faster

A new paper from Landon Liu, Mary Kelly, and Alan Tsang, accepted at CogSci 2026, explores how forgetting shapes shared meaning in groups of agents. Using a non-partnership coordination game—a departure from traditional games where players share payoffs—the researchers simulated conversations between agents with different memory characteristics: some were adaptive (forgetful, discounting old information) and others were non-adaptive (rigid, retaining all information equally).

Key findings: Adaptive agents achieved actual conceptual convergence faster and ended with closer semantic regions. But non-adaptive agents perceived convergence earlier—a false sense of agreement. Additionally, gradually reducing the weight of new information over time produced more stable agreements than keeping a constant weight. The study underscores that memory traits like forgetting and adaptiveness are critical for the emergence of shared meaning, with implications for multiagent AI systems, collaborative bots, and human-machine communication.