High-fidelity social learning via shared episodic memories enhances collaborative foraging through mnemonic convergence

TL;DR

This work investigates how episodic memory and social learning interact in collective foraging by using Sequential Episodic Control (SEC) agents that can share complete episodic memories. By manipulating memory length, transfer rate $Tr$, and transfer noise $Tn$, the study demonstrates that high-fidelity social learning consistently improves resource collection and equitable distribution, with performance scaling with the frequency of social transmissions. mnemonic metrics show that high-fidelity sharing promotes mnemonic alignment and memory distribution across agents while reducing diversity, and correlations between memory distribution and rewards are strong ($r$ around 0.93). An optimal memory length exists beyond which performance plateaus, and low-fidelity learning increases mnemonic diversity without translating into performance gains. These results illuminate how memory fidelity and distribution shape collective cognition and offer a neurocomputational lens on cultural evolution in multi-agent systems.

Abstract

Social learning, a cornerstone of cultural evolution, enables individuals to acquire knowledge by observing and imitating others. At the heart of its efficacy lies episodic memory, which encodes specific behavioral sequences to facilitate learning and decision-making. This study explores the interrelation between episodic memory and social learning in collective foraging. Using Sequential Episodic Control (SEC) agents capable of sharing complete behavioral sequences stored in episodic memory, we investigate how variations in the frequency and fidelity of social learning influence collaborative foraging performance. Furthermore, we analyze the effects of social learning on the content and distribution of episodic memories across the group. High-fidelity social learning is shown to consistently enhance resource collection efficiency and distribution, with benefits sustained across memory lengths. In contrast, low-fidelity learning fails to outperform nonsocial learning, spreading diverse but ineffective mnemonic patterns. Novel analyses using mnemonic metrics reveal that high-fidelity social learning also fosters mnemonic group alignment and equitable resource distribution, while low-fidelity conditions increase mnemonic diversity without translating to performance gains. Additionally, we identify an optimal range for episodic memory length in this task, beyond which performance plateaus. These findings underscore the critical effects of social learning on mnemonic group alignment and distribution and highlight the potential of neurocomputational models to probe the cognitive mechanisms driving cultural evolution.

Figures (5)

• Figure 1: Collective foraging task modelled in a 2D grid-world environment. The environment contains four agents, four fruits, and a nest. Fruits are represented as red circles; the nest as a green square. The agents are represented as triangles of different colors (green, yellow, blue, and purple). The 3x3 colored area around each agent represents their field of view.
• Figure 2: Panel A: Diagram of the SEC model for the 2D grid-world collective foraging task. SEC can be functionally divided into storage and retrieval phases. During the storage phase, the agent stores state-action ($s,a$) couplets into the short-term memory (STM) at each time step. When encountering a reward ($r$), the content of the STM buffer is transferred into the long-term memory (LTM). During the retrieval phase, the agent retrieves from its LTM the episodic memories more similar to the currently observed state. Based on this retrieved mnemonic information, the agent computes the state-action value function ($Q(s,a)$) for the current state and selects an action ($a$) based on the resulting probability distribution. Panel B: Social learning between two SEC agents. Agent 1 (blue) retrieves a copy of a complete episodic memory from its long-term memory (LTM) and transfers it to Agent 2 (pink), which stores the copy in its LTM.
• Figure 3: Performance results of Sequential Episodic Control agents across different social learning conditions. Top panels: results for high-fidelity social learning (Transfer Noise, Tn=0), Bottom panels: results for low-fidelity social learning (Tn=0.1). Left: average reward per episode. Center: total accumulated reward. Right: Ratio of reward distribution between agents. Colors represent the frequency of social interaction; green: no interaction (Tr=0), blue: infrequent interaction (every 50 episodes, Tr=50), royal blue: frequent interaction (every 10 episodes, Tr=10), continuous interaction (every episode, Tr=1). For clarity, only no interaction (Tr=0) and continuous interaction (Tr=1) results are shown on the left panels.
• Figure 4: Mnemonic diversity results for both high-fidelity (top) and low-fidelity (bottom) social learning conditions. Left: average relative diversity (denotes the number of unique episodic memories in an agent’s long-term memory, relative to the total amount of memories acquired by the agent). Center: group diversity (a group-level metric that captures the relative diversity of the aggregated group buffer). Right: group alignment (captures the similarity in terms of content between the long-term memories of agents within a group).
• Figure 5: Mnemonic distribution strongly correlates with reward distribution and group alignment during high-fidelity social learning. Left panels: Mnemonic distribution results. Center: Correlations between memory distribution and reward distribution. Right: Correlations between memory distribution and group alignment. The top panels show results for high-fidelity social learning conditions. The bottom panels show results for low-fidelity social learning conditions.