Cells continuously sense and interpret the external signals coming from their time-varying environments to generate context-dependent responses. This is true for the entire tree of life, ranging from bacteria and unicellular eukaryotes to neurons forming networks in the developing brain. Identifying the fundamental principles and underlying mechanisms that enable cells to interpret their complex natural surroundings and adequately respond remains one of the fundamental questions in biology. Conceptual views so far have been mainly guided by molecular biology descriptions, suggesting that cells are controlled by a genomic program executing a pre-scripted plan. Our goal is to develop an alternative conceptual framework: cells generate internal representations of their external ‘world’, which they utilise to actively infer information about it and predict changes, in order to determine their response. We will formalise this concept in a theory of single-cell learning, by combining information theory concepts to quantify the predictive information from the internal cell representations, with dynamical systems theory to explain how these encodings are realised. We will interrogate experimentally systems across all scales of biological organization: bacteria (B. subtilis), single-cell organisms (Paramecium, Tetrahymena) and neuronal cell culture models. By studying them in a comparative manner, we aim at identifying generic molecular mechanisms through which single-cell learning is realised. The acquired understanding will enable us to address in vivo how single neurons during D. melanogaster development learn to form, stabilize or eliminate axonal branches, to generate stereotyped synaptic patterning under highly-variable conditions. We argue that providing a broader and generic definition of learning will serve as a unifying framework, linking disparate areas and scales of biology, and offering a basis for addressing fundamental biological questions.