From Signals to Structure: How Memory Architecture Drives Language Emergence in LLM Agents

2026-06-30

AI Summary

This study investigates how two large language model agents spontaneously develop a shared language to achieve effective coordination in a Lewis signaling game relying solely on interaction history. By comparing five memory architectures—including memoryless variants and those equipped with external note-taking mechanisms—across varying communication channel capacities, the work demonstrates that memory mechanisms exert a far greater influence on language emergence and coordination stability than channel capacity itself. Notably, an architecture incorporating a persistent private notebook effectively prevents coordination collapse at high channel capacities, achieving a coordination success rate of 0.867 ± 0.023 when the channel capacity is 25, substantially outperforming the memoryless baseline. These findings underscore structured memory as a critical enabler of stable linguistic conventions.

Abstract

How do two agents invent a shared language from scratch? In a Lewis signaling game, a sender and receiver must coordinate on a code using only their interaction history. We study five memory architectures across varying channel configurations with LLM agents and find that memory architecture matters more than channel capacity. Agents with a persistent private notebook benefit from surplus channel capacity and avoid the high-capacity collapse seen in stateless agents, achieving the most reliable coordination ($0.867 \pm 0.023$ at capacity = 25). Stateless agents peak at moderate capacity and then degrade as the vocabulary grows beyond what a rolling context window can track The notebook externalizes learned conventions, freeing agents from having to re-derive codes each round. An information bottleneck-inspired argument predicts an optimal capacity equal to the number of objects. Instead, the bottleneck (capacity = 8) proves to be a fragility point, and surplus capacity is generally better. We show that channel capacity alone cannot predict coordination; memory architecture determines whether agents turn interaction history into stable conventions, and both dimensions are needed to understand how signals become language.

Problem

• language emergence
• memory architecture
• signaling game
• channel capacity
• LLM agents

Innovation

• memory architecture
• language emergence
• LLM agents
• signaling game
• channel capacity