LLM agents: Memory architecture shapes language emergence

From Signals to Structure: How Memory Architecture Drives Language Emergence in LLM Agents, submitted to arXiv on 30 Jun 2026, shows memory architecture matters more than channel capacity when two LLM agents invent a shared code. The authors evaluate five memory architectures in a Lewis signaling game and report the highest coordination, 0.867 ± 0.023, for agents using a persistent private notebook at capacity = 25.

What did the authors test and how?

They put pairs of LLM agents into a Lewis signaling game and varied both memory architecture and channel capacity, studying five memory architectures across varying channel configurations. The setup required a sender and a receiver to coordinate on a code using only their interaction history; the experiments compared stateless agents, agents with persistent private notebooks, and three other memory variants across different vocabulary capacities.

The paper contrasts agents that rely on a rolling context window with agents that can externalize history, and it frames capacity choices against an information-bottleneck-inspired prediction that optimal capacity equals the number of objects. Instead, the authors find capacity interacts with architecture in nontrivial ways.

How does memory architecture affect language emergence?

Memory architecture determines whether interaction history becomes stable conventions: agents with a persistent private notebook avoid the high-capacity collapse that plagues stateless agents and benefit from surplus channel capacity. The notebook externalizes learned conventions so agents do not have to re-derive codes each round, producing the most reliable coordination measured, 0.867 ± 0.023 at capacity = 25.

By contrast, stateless agents peak at moderate capacity and then degrade as the vocabulary grows beyond what a rolling context window can track. The authors identify capacity = 8 as a fragility point where the information-bottleneck-inspired argument fails: rather than matching the number of objects, that bottleneck proves fragile and surplus capacity is generally better for robust coordination.

Why it matters

This study shifts emphasis from channel capacity alone to the combination of channel capacity and memory architecture. For multiagent setups built from LLMs, the result means engineers cannot predict emergent communication solely by increasing vocabulary or bandwidth; they must choose how agents store and reference past interactions. Designing persistent private memory mechanisms can turn transient interaction traces into stable conventions and avoid collapse as signaling spaces grow.

What to watch

Observe whether other teams replicate the reported 0.867 ± 0.023 coordination for notebook-equipped agents and whether the capacity = 8 fragility point reappears in different task families. The next concrete signals will be named-architecture benchmarks showing per-architecture coordination curves across capacities and any public code or data the authors attach to the arXiv submission.

Additional context

The authors make an explicit information-theoretic link by comparing empirical results to an information bottleneck argument that predicts an optimal capacity equal to the number of objects, then show that empirical best-practices diverge: surplus capacity often helps, provided the memory architecture lets agents externalize conventions. The paper appears as arXiv:2607.00233 [cs.AI], submitted 30 Jun 2026, by Yashar Talebirad, Eden Redman, Ali Parsaee, and Osmar R. Zaiane.