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

Short summary

Researchers tested how LLM agents develop shared languages through different memory architectures in a Lewis signaling game. Agents with persistent notebooks achieved superior coordination (0.867±0.023) and avoided degradation that stateless agents experience as vocabulary grows. The findings show memory architecture matters more than channel capacity alone for converting interaction history into stable communication conventions.

• •Persistent memory (private notebooks) enables more reliable agent coordination than stateless systems
• •Memory architecture is a stronger predictor of successful communication than channel capacity
• •Surplus channel capacity helps agents with notebooks but causes degradation in stateless designs

Generated with AI, which can make mistakes.