Memory Architecture, Not Channel Capacity, Drives Language Emergence in LLM Agents

A new paper from the University of Alberta investigates how two LLM agents can invent a shared language from scratch using only their interaction history—a classic Lewis signaling game. The researchers, Talebirad, Redman, Parsaee, and Zaiane, compared five different memory architectures across varying channel capacities (the number of distinct signals available). Their counterintuitive finding: memory architecture matters more than how many signals agents can send. Agents equipped with a persistent private notebook—a dedicated external memory store that grows over time—achieved the highest coordination score of 0.867 ± 0.023 at a capacity of 25. They were able to leverage surplus channel capacity without suffering from what the authors call 'high-capacity collapse,' a degradation seen in stateless agents whose limited context window gets overwhelmed as the vocabulary grows beyond what they can track.

In contrast, stateless agents (which rely solely on rolling context windows) peaked at moderate capacities and then declined, unable to maintain consistent conventions. The theoretical prediction from an information bottleneck argument—that optimal capacity should equal the number of objects (8 in this case)—proved incorrect. Instead, that bottleneck point was a fragility threshold, and surplus capacity generally improved coordination. The notebook effectively externalizes learned conventions, freeing agents from re-deriving codes each round. The study demonstrates that channel capacity alone cannot predict coordination; memory architecture determines whether agents can move from transient signals to stable, shared language. This has implications for designing multi-agent AI systems, suggesting that persistent memory is more important than bandwidth for emergent communication.