Source-linked AI summary
Memory Is Communication: The Frontier Between Remembering and Signaling
Yashar Talebirad, Eden Redman, Ali Parsaee, Osmar R. Zaiane
TL;DR
The paper asks how bounded agents should divide information budgets between memory and peer communication. It defines and tests a remembering–signaling frontier, finding contrasting effects of target repetition and cyclic predictability on successful message length.
Problem
It remains unclear how retained history and peer communication should be allocated under joint memory and bandwidth limits at fixed task performance.
Method
The paper formalizes achievable memory–message rate regions under fixed tasks, decoders, source rules, and encoding schemes, then proposes rate-varying tests across cooperative tasks.
Results
Mean Lmin fell from 2.67 to 1.00 as target repetition increased, but rose from 2.33 to 3.00 as cyclic-sequence predictability increased.
Takeaways & Limitations
Target repetition may shift efficient allocations toward memory, whereas predictable cyclic structure did not produce shorter successful messages in the preliminary games.
Takeaways & Limitations
The preliminary evidence does not isolate history’s contribution and comes from one model family, three seeds per batch, and forty interaction rounds.
Abstract
from arXiv · showhide
A bounded agent may obtain information for a decision from its own past, from peers, or from both sources. Retaining task-relevant history can reduce later communication, while a peer message can supply what memory lacks. Under limits on both resources, how should an agent allocate its information budget? Given a fixed task and decision rule, the memory and message rate pairs attaining a performance threshold form an achievable region under specified rules for using history and peer observations. We call its efficient boundary the remembering--signaling frontier. Across conditions where history permits the same maximum reduction in task loss, we hypothesize that a bounded agent will need less peer communication when it obtains a larger loss reduction from history. In preliminary referential games, target repetition coincided with shorter successful messages, while predictability from a hidden cyclic rule did not shorten them. Experiments varying memory and message rates can estimate the frontier and test this prediction across cooperative tasks.
Introduction
The section frames cooperative decision-making as an information-allocation problem for agents with partial observations and bounded memory and communication. It defines “memory is communication” as retained history replacing peer input without changing task performance.
- Introduction: Cooperative agents act from partial observations while constrained by memory and communication bandwidth.Existing protocols optimize which agents communicate, what they send, and to whom.
- Introduction: “Memory is communication” describes retained history replacing peer input at fixed task performance.The framing treats memory as communication between temporally separated selves.
Framework and Hypotheses
The framework defines an achievable memory–communication region for a fixed task, decoder, source rules, and scheme family, with its Pareto-efficient lower boundary called the remembering–signaling frontier. It tests whether usable-history gain predicts reduced peer communication and whether one-rate-at-a-time loss curves identify the least-cost joint allocation.
- Framework: Source rules specify whose past records and peer observations may be encoded, when they are available, and exclude future or otherwise unavailable information.The task includes its instance distribution and loss, while decoder A maps current observations and supplied representations to the output.
- Framework: The achievable region contains memory–communication budget pairs whose allowed schemes attain expected loss at most ε, and its Pareto-efficient lower boundary is the remembering–signaling frontier.The region depends on the task, decoder, source rules, and scheme family.
- Hypotheses: The formulation counts both historical-representation bits and peer-message bits, asking whether history’s loss reduction predicts lower message rates at the same task loss.Memory and communication are treated as jointly constrained information resources.
- Evaluation: Usable-history gain is the loss reduction from adding history, while usable-peer gain is the additional reduction from adding peer input under fixed rates, source order, and loss measure.Evaluation uses held-out instances and three conditions: current observation alone, history added, and peer message added.
- Hypotheses: The second hypothesis predicts that varying memory and message rates separately can identify the least-cost allocation when both rates vary jointly.Loss is first measured as memory rate increases without peer messages, then as message rate increases with memory fixed in advance.
Preliminary Results
Preliminary signaling-game results linked target repetition to shorter successful messages, whereas predictability from a hidden cyclic rule produced longer messages. The repetition runs did not isolate history’s contribution because episode history and learned symbol mappings were retained.
- Repeat condition: Mean Lmin fell monotonically from 2.67 at p = 0 to 1.00 at p = 0.95 when targets repeated probabilistically.The repeat condition copied the previous target with probability p and otherwise sampled uniformly, across two batches of three seeds.
- Interpretation: The repeat-condition runs did not isolate history’s contribution because episode history was never removed and learned symbol mappings were preserved.Thus, shorter messages could reflect retained history, learned mappings, or both.
Discussion
The discussion identifies limitations in the preliminary evidence and proposes three progressively general tests of the remembering–signaling frontier. These tests span exact state estimation, controlled referential-game rate experiments, and distributed coordination tasks with local views.
- Limitations: The preliminary hypothesis remains untested because usable-history measurements and message-rate predictions have not been preregistered.Interpretation is also limited by differences in rule complexity and target counts, one model family, three seeds per batch, and forty interaction rounds.
- Three tests: The first test computes the Bayes-optimal frontier exactly in a finite noisy state-estimation task combining compressed self-history with rate-limited peer messages.The hidden state persists with fixed probability or changes, allowing comparison with a learned decoder’s frontier.
- Three tests: The second varies memory rate, message rate, and target predictability in referential games while matching target frequencies and Bayes-optimal historical gains across processes.A memory ablation removes episode history while preserving learned symbol mappings; longer runs and another model family test learning-time and capability explanations.
- Three tests: The third repeats rate experiments on distributed local-view tasks, varying the persistence of private inputs or local states while holding the communication graph fixed.The preregistered subset includes local and global coordination tasks drawn from AgentsNet and LoopBench.