Accurate tool wear monitoring is crucial for the efficient and stable operation of smart manufacturing systems. However, multi-source sensor signals in complex cutting environments often couple high-frequency transient impacts with low-frequency long-term degradation trends. Traditional deep learning models, such as Long Short-Term Memory (LSTM) networks, frequently encounter challenges regarding long-range historical information decay and localized perceptual imbalance when processing such multi-scale features. To address these bottlenecks, this paper proposes a Bidirectional Long-term Attentive State Space Model (Bi-LASM). This model integrates bidirectional recurrent gating mechanisms with Selective State Space Models. Specifically, an adaptive multi-source signal fusion mechanism is designed to extract condition-sensitive channel features. Subsequently, an input-dependent dual-scale linear state space branch is constructed to decouple and parallelly fit the long-term smooth degradation patterns and short-term local mutation characteristics of the tool. Furthermore, a dynamic gating mechanism is introduced to achieve step-level adaptive weight allocation between long-term and short-term memory. Comprehensive experiments on an actual impeller milling dataset integrating force, vibration, sound pressure, and current sensors demonstrate that Bi-LASM achieves an overall recognition accuracy of 98.12%. This performance yields a 1.87% improvement over the pure Selective SSM baseline. Furthermore, the proposed model consistently outperforms mainstream architectures, including the Transformer and Bi-LSTM. The proposed method exhibits stable performance under complex working conditions and provides explicit physical interpretability for tool state evolution.