A Computational Framework for Social Modulation of Generative Episodic Memory

Episodic recall is reconstructive, combining incomplete episodic information with learned semantic knowledge. In social settings, reconstruction can additionally be influenced by information acquired from other individuals, producing systematic biases and distortions in memory. Although such effects are well established experimentally, computational accounts that jointly formalise episodic, semantic, and socially acquired information remain limited. We extend a generative framework of episodic memory through substantial revisions to its architectural components, parameter regimes, and optimisation objectives, improving its biological plausibility while introducing social information as an additional source of reconstructive evidence. Visual episodes are encoded as discrete latent memory traces using a Vector Quantized-Variational Auto Encoder (VQ-VAE) and partially masked to reflect the inherent limitations of human attention. They are then completed by a bidirectional masked transformer using retained episodic content, learned semantic regularities, and socially supplied context. We formalise social influence through two complementary routes: (i) forming additional traces during social interaction that can persist and shape later recall, and (ii) biasing semantic completion during recall through social context. We evaluate the framework across three paradigms of social influence on memory, namely the saying-is-believing paradigm, co-witness memory conformity, and social contagion through repeated exposure to misinformation. Specific experimental conditions are instantiated within our framework, which reproduce characteristic behavioural patterns reported in experimental results of corresponding human studies. For instance, audience-congruent recall increases under shared-reality conditions while remains unchanged under non-shared reality conditions, mistaken recall of details acquired only through co-witness discussion, and stronger effects of repeated misinformation. The framework thereby provides a common computational basis for simulating diverse social-memory paradigms and systematically investigating how episodic, semantic, and socially acquired information jointly shape reconstructive recall. The source code for our framework is available at https://anonymous.4open.science/r/socialgemmodulation-3C0F/ .