Long-term effects of the Severance Schema under accumulating user memory

Determine the long-term effects of incorporating the Severance Schema—a prompt-level structured awareness scaffold that enumerates unknown dimensions of person-context—into personalized AI assistants as user-specific memory continuously accumulates across sessions.

Background

The paper introduces the Severance Problem: LLMs lack both information about the user and an explicit representation of what categories of person-context they are missing. To address this, the authors propose the Severance Schema, which organizes user context into six dimensions (physicality, temporality, consequences, continuity, multiplicity, and interiority) and explicitly marks unknowns.

Experiments on synthetic advisory scenarios show that adding the Severance Schema improves unknown-awareness, reduces sycophancy and harmful advice, and mitigates hallucinations, both with and without memory. However, these evaluations focus on short-term interactions. The authors explicitly note that the long-term effects of deploying the Severance Schema as user memory accumulates remain unresolved and plan to investigate this via IRB-approved user studies.