Algorithm 1 Modulation Context Memory Module

Input: fused tokens ${\mathbf{Z}}_0\in {\mathbb{R}}^{B\times T\times D}$. Output: refined tokens ${\mathbf{Z}}_m\in {\mathbb{R}}^{B\times T\times D_m}$. Channel–temporal recalibration:     Switch fused tokens to the channel-first layout: $\mathbf{U}={\mathbf{Z}}_0^{\top }$.     Enhance energy-related modulation responses: $\mathbf{E}=\mathbf{U}\odot \mathbf{U}$.     Encode channel–temporal dependency: ${\mathbf{H}}_a={\mathrm{GRU}}_a\left({\mathbf{E}}^{\top }\right)$.     Generate the recalibration gate: $\mathbf{G}=\sigma {({\mathbf{W}}_a{\mathbf{H}}_a+{\mathbf{b}}_a)}^{\top }$. Preserve residual evidence: ${\mathbf{U}}_a=\mathbf{U}\odot \mathbf{G}+\mathbf{U}$. Bidirectional temporal memory:     Aggregate past and future context: $({\overrightarrow{\mathbf{H}}}_t,{\overleftarrow{\mathbf{H}}}_t)=\mathrm{BiLSTM}\left({\mathbf{U}}_a^{\top }\right)$.     Form the refined token sequence: ${\mathbf{Z}}_m=[{\overrightarrow{\mathbf{H}}}_t;{\overleftarrow{\mathbf{H}}}_t]$. return ${\mathbf{Z}}_m$.