Identity-Conditioned Forgetting in Episodic Agent Memory

We introduce identity-conditioned forgetting, a memory mechanism in which past experiences slowly fade in feature space over time. Importantly, how quickly an experience fades depends on how closely it matches the agent’s internal identity. Memory is represented as a time-ordered sequence of events, where each event is encoded using CNN feature representations and records outcome-specific traces through separate success and failure branches. Forgetting is implemented as a continuous degradation of these representations, with memory retention controlled by cosine alignment to a latent identity prior (self.facts).