Functional Scars: A Runtime Mechanism for Persistent Corrections in AI Agent Systems

Abstract

An AI agent can retain a correction across sessions and still repeat the corrected action. A record may be missing, remain unretrieved, arrive after the decision point, or be read without changing the action. This paper defines a functional scar as a failure-derived runtime control that links a documented error to an applicability predicate, an observable obligation, an intervention point, an evidence record, and a review path. The definition separates the persistence of information from the persistence of corrected behavior. It gives a platform-neutral contract for representing and testing such controls, then examines an open-source implementation with Claude Code and Codex adapters. In the DG Office deployment, Claude Code runs a locally maintained hook system, while Codex invokes the fscars package through native hooks; the two arrangements share a control idea and differ in implementation. A reproduced 368-test suite establishes conformance to the tested software contracts. Prior practitioner observations report lower recurrence for some structured corrections. Runtime logs use different event schemas and outcome proxies, so they do not estimate a package-wide causal effect. The paper sets out an evaluation method that measures applicability, control activation, action change, final outcome, false activation, and override cost as separate quantities. A brief analogy with human emotional learning motivates the term, while neuroscience is kept distinct from the engineering mechanism. This technical synthesis specifies a testable correction-control contract and evidence boundaries for evaluating persistent corrections.