Lucy Syndrome in LLM Agents: A Practitioner Framework for Cross-Session Correction Persistence

Large Language Model agents operating in production environments systematically lose corrections between sessions. An operator fixes an error, the model acknowledges the fix, and the next session reproduces the original mistake as if the correction never happened. This paper names the pattern "Lucy Syndrome" and proposes a practitioner framework for making corrections persist. The study is based on a structured lab conducted over the author's own production logs from a civil engineering firm operating nine business areas through Claude Code. From months of operational data, 163 findings were extracted and cross-referenced across four analytical categories: repeated errors, successful corrections, false confidences, and metacognitive events. The cross-reference reveals that the four categories are not independent phenomena but a single causal chain. The central contribution is a framework of five invariants that distinguish corrections that persist from those that decay, operationalized through "functional scars" — hook-based enforcement mechanisms that convert each operator correction into an automated check the model cannot skip.