Explicit Retrieval, Implicit Cognition: Towards Personalized Micro-video Popularity Prediction via Dual Latent Memory

Micro-video popularity prediction (MVPP) is crucial for product marketing and recommendation. Existing works typically predict popularity by using independent modality encoders, but are limited by a personalized reasoning bottleneck: a tendency to lose grounding in multimodal evidence and a lack of personalized experience during reasoning. Inspired by human memory theory, we propose CLAMP, a cognition-memory-aligned framework that synergizes a large multimodal model (LMM) with dual latent memory, comprising a short-term module for content-centric perceptual retention and a long-term module for user-centric semantic consolidation. To achieve personalization, CLAMP develops a flexible memory retrieval mechanism governed by a familiarity-recollection process: familiarity enables fast recognition, while recollection supports deliberate, chain-like episodic retrieval. This mechanism adaptively switches between dual retrieval pathways to search for relevant memory knowledge. Furthermore, to invoke memory, CLAMP designs a dynamic invocation module that adaptively generates latent memory tokens to retain both perceptual evidence and user-level semantics. These memories are seamlessly invoked during inference, allowing the LMM to maintain perceptual fidelity and personalized semantics throughout both thinking and generation. Extensive experiments show that CLAMP consistently outperforms all baselines, validating its efficacy in personalized MVPP reasoning.