Valuation Is Authored, Not Observed: Separating Fact, Affect, Norm and Narrative in Personal Memory Substrates for AI Agents

Memory systems for AI agents increasingly store how a person felt, what they value and what their life means, usually in the same store as facts and usually written by the system rather than the person. This paper proposes a plane model for personal memory substrates with five typed planes (Evidence, Fact, Appraisal, Norm, Narrative). Each record carries an authorship class (stated, confirmed, inferred or institutional), and six write-time invariants govern derivation, promotion and release. The central principle is that valuation is authored, not observed: inferred feelings, norms or life lessons are hypotheses that may not be promoted, exported or used to gate action without the person’s confirmation. The paper extends Minimum Viable Exposure with plane-dependent exposure floors and maps the design to the GDPR (CJEU C-184/20) and the EU AI Act. From that mapping it derives a workplace profile with no affect plane. A reference SQLite schema and a reproducible worked example are included. Design and position paper; no user study.