Cross-domain sequential recommendation (CDSR) leverages user behaviors across domains to alleviate data sparsity and enhance next-item prediction. However, enriching item semantics with tags, textual descriptions, or LLM-derived knowledge may inadvertently introduce negative transfer. We identify two primary causes: (1) semantic dominance over collaborative signals, which overlooks the domain-specific semantics underlying identical labels, and (2) symmetric fusion of global features, which ignores the inherently asymmetric nature of cross-domain sharing. To tackle these issues, we propose a Collaborative Memory and Gradient-Feedback Reweighting model (CMGR). CMGR constructs a global semantic memory by clustering LLM-derived item embeddings into semantic prototypes and querying them with mixed-domain user sequences to sparsely retrieve shared semantics. By injecting user-item collaborative signals for calibration, the memory module resolves domain-specific semantic ambiguity and enhances alignment for cold-start items. Furthermore, CMGR introduces a gradient-feedback reweighting mechanism that treats backpropagation gradients as implicit supervision to estimate the instantaneous marginal utility of shared features and dynamically adjust their domain-wise contributions. Extensive experiments on four real-world CDSR datasets demonstrate that CMGR consistently outperforms state-of-the-art methods.