Agglutinative languages encode a substantial part of grammatical relations through ordered affix sequences attached to lexical stems. Encoders in the BERT family can learn some of these regularities implicitly from large corpora, but conventional self-attention does not treat lemma, stem/root, part of speech, affixes, or grammatical categories as separate information sources. This study proposes a new model named Morph-QKV. The model builds a dedicated morphology representation M, derives morphological Query, Key, and Value projections from it, and extends the native attention score with morphology-to-morphology, contentto-morphology, and morphology-to-content interactions. Morph-QKV is developed in architecture-specific mathematical forms for absolute-position BERT, instantiated with Tahrirchi-BERT; disentangled relative-position DeBERTa, instantiated with mDeBERTa-v3-base; and RoPE-based local/global ModernBERT, instantiated with mmBERT-base. Results from MAGNet, LISA, LUKE, tokenizer studies, and character/byte-level encoders demonstrate the value of explicit linguistic signals. On this basis, Morph-QKV is designed to improve accuracy and Macro-F1 on Uzbek NLP tasks.