Machine-learning systems can encode molecular structures, predict chemical transformations, generate candidate compounds, and infer biomedical relationships, yet these capabilities are commonly distributed across models that do not share a coherent scientific reasoning framework. Learned representations can identify statistical regularities without preserving explicit chemical syntax, reaction applicability, biological context, or the evidentiary status of mechanistic claims. Conversely, symbolic systems can express rules and ontological relationships but may be brittle, incomplete, and poorly adapted to uncertain or previously unseen molecular contexts. This article proposes a neuro-symbolic molecular architecture designed to reconnect these representational modes through bidirectional reasoning. The proposed system contains learned molecular and reaction encoders, explicit chemical and biological knowledge layers, neural-to-symbolic grounding, symbolic-to-neural feedback, constraint adjudication, uncertainty estimation, contradiction preservation, provenance tracking, and task-specific reasoning interfaces. Its central contribution is not a new predictive score but an architectural account of how molecular hypotheses could be generated, checked, revised, qualified, or deferred through interactions among embeddings, reaction rules, mechanistic knowledge, and scientific constraints. The framework distinguishes syntactic validity from reaction feasibility, association from mechanism, model confidence from calibrated uncertainty, and benchmark performance from prospective pharmaceutical usefulness. It also specifies how the architecture should be evaluated separately in molecular design, synthesis planning, and mechanism inference. The proposed system remains conceptual: it has not been empirically validated, does not establish causal or therapeutic claims, and cannot substitute for experimental confirmation, medicinal-chemistry judgement, or regulatory assessment. Its intended value is to provide a disciplined foundation for developing molecular AI systems whose outputs are not merely plausible or fluent, but scientifically inspectable, constraint-aware, and appropriately bounded.