Memory Is Communication: The Frontier Between Remembering and Signaling

In the world of intelligent systems, whether biological or artificial, there is a constant tension between what an entity remembers and what it must ask others. Imagine a team of explorers trying to navigate a dark forest. Each explorer has a limited capacity to carry a map in their head, and they have a limited ability to shout instructions to their teammates. If an explorer remembers the path they just took, they might not need to ask for directions. Conversely, if a teammate shouts a clear warning, the first explorer might not need to remember as much. This trade-off is the core of a new line of research exploring how agents—be they computer programs or robots—share the burden of knowing. The question is not just about how much information can be stored or sent, but how to split the cost between keeping a history of the past and signaling the present. When resources are tight, the most efficient strategy is not obvious, and finding the perfect balance could help us build smarter, more cooperative machines.

Researchers at the University of Alberta and the Network for Applied Technology have begun to map this balance, proposing a concept they call the "remembering–signaling frontier." This frontier represents the most efficient way to combine memory and communication to solve a task. If an agent uses more memory to recall past events, it can afford to send fewer messages. If it relies less on memory, it must send more messages to achieve the same result. The researchers hypothesize that if an agent can extract a lot of useful information from its own history, it will need significantly less help from its peers. To test this, they set up a series of experiments where digital agents played a game of pointing and guessing, trying to identify a specific object among many options while strictly limiting how much data they could store or transmit.

The experiments took place in a simplified version of a signaling game. In each round, a "sender" agent looked at four shapes on a screen, each with a unique combination of color, size, and form. One of these shapes was the target. The sender had to communicate this target to a "receiver" agent using a very short code made of symbols like A, B, or C. The receiver then had to guess the correct shape. The catch was that the agents could only send a limited number of symbols, and they had to rely on a private notebook to remember their past interactions. The researchers wanted to see how the agents adapted their strategy when the target became more predictable. In one scenario, the target often repeated the same shape from the previous round. In another, the targets followed a hidden, rotating cycle that the agents could not see, but which repeated over time.

The results revealed a fascinating difference in how the agents handled these two types of predictability. When the target simply repeated the previous one, the agents achieved success with fewer symbols as the probability of repetition increased. At the highest level of repetition, they could almost always identify the target with just a single symbol. Because the agents retained their private notebooks and preserved their learned symbol mappings throughout the runs, these results did not isolate the specific contribution of history alone. However, the situation changed dramatically when the targets followed the hidden rotating cycle. Even though the pattern was perfectly regular and predictable in theory, the agents did not shorten their messages as the rotation became more consistent; in fact, they needed to send longer messages to succeed. The interpretation is limited because the target processes differed in rule complexity and in the number of distinct targets encountered, rather than solely due to an inability to use memory.

This contrast suggests that not all history is created equal. The researchers found that the ability to reduce communication costs depends heavily on the nature of the information stored. When the past directly predicts the present in a simple way, memory acts as a powerful substitute for communication. But when the pattern is complex or hidden, simply remembering the past does not automatically translate into shorter messages. The team suspects that the agents were unable to compress the rotating pattern into a useful memory format, forcing them to rely more heavily on sending fresh information. These early findings, derived from simulations using large language models, are not yet a final proof. The researchers acknowledge that their initial tests were limited by the specific models used and the short duration of the games.

To move beyond these preliminary hints, the team is designing more rigorous tests. They plan to use simpler, mathematically solvable tasks where the perfect answer is known, allowing them to compare what the learning agents achieve against the theoretical best. They also intend to expand their experiments to more complex group tasks, such as coordinating a network of agents to solve problems like coloring a map or electing a leader. By systematically varying how much memory is available and how much communication is allowed, they hope to draw the full "remembering–signaling frontier" for different types of problems. The goal is to understand exactly when an agent should look inward to its own history and when it should look outward to its peers. Until then, the study stands as a careful exploration of the boundary between what we keep in our minds and what we must say to one another.