Forgetting in the Machine: Semantic Compaction and Fixation in Log-Oriented Systems

Modern information systems increasingly adopt the log as their primary storage abstraction. This creates a fundamental tension: logs grow without bound, while resources remain finite. Compaction mechanisms—such as log compaction, snapshotting, and retention policies—are employed to reconcile this tension by reducing historical data. This paper argues that compaction is not merely a storage optimization but a semantic operation of forgetting, inseparable from the interpretation of events. We prove that no universal compaction operator exists independently of semantic interpretation. As a consequence, every compaction policy implicitly defines which historical distinctions remain meaningful for system objects. We introduce the concept of fixation—the process that determines which distinctions in history are constitutive of an object’s identity. Analyzing the architectures of distributed systems, databases, and applications, we show that choosing a fixation strategy is always choosing what the system is willing to sacrifice. We propose a classification of forgetting strategies and discuss their engineering and epistemological implications.