Archives are not passive storehouses; they are epistemic infrastructures that decide what can be found, trusted, cited, reused and remembered. Mbembe shows that archives transform selected fragments into public proof, while Stoler argues that archives must be studied as processes of knowledge production rather than merely mined as sources. This matters for science because datasets, specimens, metadata, citations and algorithmic records acquire authority only through institutional systems of description, preservation and access. Digital repositories such as DANS demonstrate that open data requires continuing human labour: archivists curate files, repair metadata, mediate access and make data intelligible beyond its original context. Likewise, metadata quality determines whether public data are genuinely reusable or merely nominally open. Yet archives also exclude: they privilege what is digitised, standardised and visible, while marginalising knowledge outside dominant infrastructures. Scientific archives must therefore preserve not only polished results, but uncertainty, error, context and provenance. Ultimately, trustworthy science depends on just archives: transparent, sustainable and critically aware systems that make evidence durable without pretending that memory is complete.