New Research Shows How an Agent's Memory Architecture Shapes the Language It Invents

When two people who share no common language are dropped into a room and told to cooperate, they improvise. They point, they grunt, they settle on rough conventions, and within minutes a private code begins to take shape. A new paper asks whether large language models do the same thing when paired off with no shared vocabulary, and it lands on a conclusion that should reshape how engineers think about multi-agent systems: what determines whether a durable "language" emerges is not the size of the communication channel between agents, but the way each agent is allowed to remember.

Why language emergence matters

As AI systems increasingly work in teams, autonomous agents coordinating on tasks without a human dictating every step, they need to communicate. Today that mostly happens through hand-designed protocols and natural language. But researchers studying emergent communication are interested in a deeper question: given only a channel and a shared goal, will agents invent their own structured signaling system from scratch? The answer bears directly on interpretability. A language that agents invent on their own may be efficient, but it can also drift into codes their human overseers cannot read.

The new study, "From Signals to Structure: How Memory Architecture Drives Language Emergence in LLM Agents," posted to arXiv on June 30, 2026 and accepted to the ALIFE 2026 conference, comes from Yashar Talebirad, Eden Redman, Ali Parsaee, and Osmar R. Zaïane, affiliated with the Alberta Machine Intelligence Institute at the University of Alberta and the Network for Applied Technology in Edmonton.

The method, in plain terms

The authors use a classic testbed called the Lewis signaling game. A sender sees a designated target object among a set of candidates and transmits a constrained signal; a receiver sees the same candidates plus the signal, then guesses the target. No meanings are pre-assigned. The agents converge, if they converge at all, purely through repeated rounds of trial and error. Sustained success means the pair has effectively invented a shared language. Both roles were played by the same base model, gpt-5.4-mini.

The clever move is to hold the model fixed and vary two things independently. The first is channel capacity, the number of distinct messages available, which the authors swept from 4 up to 125 by changing vocabulary size and message length. The second, and the paper's real subject, is memory architecture: what each agent retains between rounds and in what form. Every condition shares a rolling window of the last 20 interactions, but they differ in the persistent store layered on top. The five variants are: memory_only (the rolling window alone, effectively stateless); env_board (a public convention table the environment compiles and both agents read); scratchpad (a private free-form notebook of up to 150 words, rewritten each round); codebook (a fixed 10-slot list edited one operation at a time); and codebook_meta (the slot list plus a persistent meta-note). Crucially, each private store is visible only to its owning agent, and the only thing that ever crosses between them is the message itself.

What they found

The headline result is that memory architecture matters more than channel capacity. Agents given a persistent private notebook, the scratchpad condition, benefited from extra channel capacity and avoided the collapse that struck stateless agents when the code space grew large. The notebook, the authors explain, externalizes learned conventions, so agents don't have to re-derive their code from scratch each round. The most reliable coordination the team recorded was 0.867 (plus or minus 0.023) at a capacity of 25, well above chance.

Stateless agents told a different story. They peaked at moderate capacity and then degraded as the vocabulary outgrew what a rolling context window could track. A tempting information-theory prediction, that the optimal capacity should equal the number of objects, turned out to be wrong. The point where capacity exactly matched the object count (a capacity of 8) was not a sweet spot but a fragility point where outcomes split into success or failure. Surplus capacity, counterintuitively, was generally better.

The learning dynamics were revealing too. The public shared board converged fast but by copying an environment-supplied table rather than forming genuine conventions. The scratchpad showed the steepest mid-game rise but drooped in late rounds with widening variance, a sign of convention drift. The plain rolling-window condition was the most stable late in the game. The rigid codebook variants showed high variance throughout with no sustained improvement.

The authors summarize the stakes directly: "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."

Implications and what to watch

For anyone building multi-agent AI, the practical lesson is that a bigger or richer communication channel is not a substitute for giving agents the right way to remember. The form of memory, an editable notebook versus a rigid slot list versus a shared board, changes not just how well agents coordinate but what kind of conventions they form and how stable those conventions stay.

That has an interpretability edge. If a scratchpad lets agents write down and carry forward their own codes, those written conventions are a window into the private language forming between them, but the observed drift is a warning that such codes are not fixed and can quietly mutate over time. The study is deliberately minimal, two agents, one base model, a toy signaling game, so the open questions are whether these patterns hold with larger teams, heterogeneous models, and richer tasks. As agent ecosystems scale, the design of memory may prove to be one of the quiet levers that decides whether their emergent languages stay legible to us.