Instance-Dependent Interpretive Drift in Scaffolded Memory Systems: An Exploratory Study of External Memory Effects on Large Language Model Inference

External memory systems are increasingly used to compensate for the stateless behavior of large language model (LLM) sessions. Most evaluation of such systems focuses on retrieval accuracy, personalization, and task continuity. This exploratory report examines a different question: whether external memory scaffolds change the way independent LLM instances interpret the same source material. The study uses a 216-page human–AI collaboration transcript, a reflective paper written from within that collaboration, and subsequent interpretations generated under different model-state conditions. The examined conditions include a fresh instance with no prior context, an in-session instance shaped by an extended conversation, and memory-scaffolded instances initialized with curated external context such as Mempalace, Memory Chain records, and provenance artifacts. The central observation is that shared source material did not produce a single stable interpretation. Instead, instances clustered into distinguishable interpretive regimes: fresh instances tended toward abstraction and methodological distance; in-session instances tended toward relational and narrative alignment; and memory-scaffolded instances tended toward continuity preservation and coherence reinforcement. This report proposes the terms interpretive drift, coherence pressure, and verification anchoring to describe these effects. The work does not claim to establish AI consciousness, agency, or identity continuity. It treats LLM systems as research instruments and subjects of comparison, and argues that memory-augmented AI systems should be evaluated not only for what they retrieve, but for how memory scaffolds reshape inference.