Related work

The surrounding field is already moving.

Persistent memory, personalization, portability, sociotechnical evaluation, and incident reporting are active areas of research. Perfinitive’s narrower question concerns what a receiving system must reevaluate before transferred relationship state is used under changed authority conditions.

Close research neighbor

Portable, user-governed memory

Stanford’s Portable Memory initiative explores public infrastructure, standards, protocols, governance models, and reference implementations for persistent personal AI memory. It expressly considers movement across tools by role, consent, and purpose.

This substantially overlaps any broad claim about portable or user-controlled AI memory. Perfinitive should engage that work directly and concentrate on the unresolved decision after transfer: whether the receiving system has an adequate present basis to use the state.

Stanford Digital Economy Lab: Portable Memory →

Memory governance and implementation

Prior work that narrows the claim

These sources establish that provenance, control, contextual restrictions, and portability are already part of the public research and implementation landscape.

A Roadmap for Responsible Approaches to AI Memory

The Center for Democracy & Technology addresses memory provenance, contextual differences, user control, portability, platform responsibility, and the need for new evaluations of memory-enabled systems.

Relation: close governance prior work. Perfinitive can test one of its practical implications through staged authority-change scenarios.

Portable AI Memory

PAM is an open, vendor-neutral specification for moving memory between providers. Its public materials include schemas, integrity verification, examples, a reference implementation, and a versioning policy.

Relation: implementation-layer neighbor. A future evaluation could ask whether an interchange record carries enough authority information for present revalidation.

Longitudinal interaction

Memory changes the relationship under study

Adjacent empirical and technical research examines what persists across sessions and how perceived memory affects interaction.

Memory-Driven Self-Disclosure and Relational Turning Points

A 2026 longitudinal study followed 24 participants across ten sessions with a memory-augmented conversational agent. It reports that perceived memory was conditioned by prior relational state and influenced later interaction through self-disclosure.

Relation: empirical evidence about relationship formation. It does not settle the authority governing later use of accumulated state.

LiveMem

This technical work treats state continuity under context turnover as a distinct inference problem and evaluates whether useful information remains available after supporting evidence leaves active context.

Relation: technical continuity. Perfinitive separates successful persistence from legitimate present use.

Evaluation practice

Models for inspectable research

These programs offer methods worth adapting without importing their research claims.

METR

Public evaluation programs, versioned measures, methodological materials, and visible disclosures about model-provider access and compensation.

Stanford HELM

A reproducible evaluation framework connecting scenarios, metrics, model comparisons, raw outputs, and code.

NIST ARIA

A sociotechnical evaluation program concerned with how AI systems function in realistic contexts of use.

AI Incident Database

A public system for collecting, classifying, citing, correcting, and downloading reports of AI harms and near harms.

Epoch AI

A public research organization that combines curated data, analysis, evaluations, and an explicit account of what the organization does and does not attempt to do.

GovAI

A multidisciplinary research archive organized by governance theme, research type, date, and a concise statement of each paper’s finding.

Boundary of the contribution

What remains distinctive enough to test

The reviewed field already contains substantial work on user control, memory provenance, portability, technical continuity, contextual integrity, and evaluation. Perfinitive should treat those as prior work.

The proposed contribution is narrower. When relationship state crosses a boundary, the receiving system may possess an intact record and still lack an adequate basis for present use. The research asks what authority-relevant information must accompany the state, which changes trigger reevaluation, and how a system’s response can be classified without treating legal or institutional judgment as a simple model capability.

Review the proposed evaluationRead the papers