In fields that study complex systems (like the cognitive sciences), it is common to consider phenomena at distinct levels (Anderson, 1972; Love, 2012; Oppenheim & Putnam, 1958). For example, a well-known approach in cognitive science is to separate computational, algorithmic, and implementational levels of analysis (Marr, 1982; Marr & Poggio, 1976), and it is common in neuroscience to distinguish levels of organization (e.g., cellular, circuit, and systems) (Churchland & Sejnowski, 1988, Churchland/Sejnowski, 1994) and levels of mechanistic (Bechtel, 1994) or causal (Craver, 2007) explanation. However, it has been difficult to rigorously identify a single set of distinct levels based on phenomena themselves that correspond to the ways this approach is used in practice (Love, 2012, Potochnik, 2021, but see Machta et al., 2013). Rather than statements about discretization in the world, these level-based approaches can be seen as a useful problem-solving strategy (Levenstein et al., 2023), which balances the need to focus on a subset of aspects of phenomena (in order to effectively solve problems in a causally complex world), with the need for a limited number of shared abstractions (rather than a large number of problem-specific abstractions), and is especially useful in the study of biological systems or those studied under a framework of computation (for which the abstraction of processes and functions is a critical strategy, Colburn & Shute, 2007; Wouters, 2003). In addition to distinct levels, multiple lines of work have emphasized the need for accounts that unify abstractions at different levels into multi-level theories (Craver, 2002b, Bernston & Norman, 2021, O’Malley et al., 2014). In neuroscience, this often involves a division of labor into descriptive explanations (which idealize a phenomenon at a given level of abstraction), mechanistic explanations (which explain how idealizations at one level emerge from those at lower levels), and normative explanations (which explain idealizations at one level by appealing to their ability to perform a function at a higher level) (Levenstein et al., 2023). For example, the solution to the problem in Box 1 is a normative explanation in which a phenomenon (selective memory consolidation) is explained by appealing to its optimality to perform some task (memory performance in generalizable environments, as modeled at a higher level of abstraction). In addition, CLS/SCT contains descriptive explanations that idealize, e.g., hippocampal and neocortical processes, memories, and other components of the theory at, e.g., behavioral, circuit, and computational levels of abstraction, which are connected by mechanistic explanations for *how* memories are initially formed (Nadel et al., 2012) and consolidated (Klinzing et al., 2019) and normative explanations for *why* memories should be separated into complementary learning systems (Roxin & Fusi, 2013). This multi-level theory surrounds a “core” (Lakatos, 1970) idea: that the mammalian brain has two complementary learning systems and the offline replay of hippocampally stored memories supports their integration into cortical circuits. However, without the multi-level “belt,” the emperor has no proverbial clothes—it cannot make any predictions, and even the terms in the core are meaningless. The Marrian level scheme further specifies analysis at computational, algorithmic, and implementational levels. In this case, specific cognitive phenomena can be characterized in terms of the computations they enable (goals/task), the algorithms by which they do those computations (without referring to the specific implementation), and the ways in which they are implemented in neural (or non-neural!) substrates. This level scheme is especially useful for theories in the field of cognitive science (which seeks to build a useful body of knowledge about cognitive phenomena) in which the explanatory target is a specific computation/cognitive phenomenon, due to multiple realizability (Marr, 1982)—that is, the same computation could be performed by many different algorithms, and the same algorithm could be implemented in different substrates. In contrast, for theories in which the target is a neuroscientific phenomenon (i.e., in the field of neuroscience), the descriptive/mechanistic/normative division allows researchers to (1) characterize neural phenomena at a variety of levels of abstraction; (2) understand how they emerge and how they operate in ways that facilitate understanding, prediction, and possibly informed/effective manipulation; and (3) understand the functions that they serve for the operation of the brain and behavior and how those functions may be enhanced or disrupted (e.g., in disease). This flexibility is necessary because, in addition to multiple realizability, neuroscience faces an issue of multi-functionality—the same neural phenomenon can play a role in multiple functions (e.g., it does not make sense to think about the singular function of inhibitory neurons, but rather their role in many different functions, from the computation of contextual modulation, Keller et al., 2020, to maintaining a stable balanced state in neural populations, Sadeh & Clopath, 2021), many phenomena are subject of study long before a connection to specific functions are established, and there is disagreement as to what the relevant functions even are (Buzsáki, 2020, Poppel & Adolfi, 2020) or if they can all be well explained as computation (Brette, 2018; Marder & Goaillard, 2006; Richards, 2018).