Skip to content
Research hypothesis

Relationship Memory

A research hypothesis for memory that preserves useful continuity without turning every interaction into a permanent record.

Continuity with boundaries

Relationship memory should help a system remember stable preferences, prior commitments, and the meaning of shared events. It should not retain everything by default or silently transform uncertain interpretations into permanent facts.

Proposed memory layers

We distinguish session state, user-approved preferences, episodic summaries, and revisable relationship hypotheses. Every layer needs provenance, scope, a way to correct it, and a way to delete it.

Evidence boundary

Prior work shows that stored experiences, reflection, and retrieval can support coherent agent behavior. Our privacy-preserving relationship model is a separate hypothesis and has not yet been validated through longitudinal user research.

Sources and provenance

  1. Park et al., Generative Agents (UIST 2023)
  2. AHI, companion memory experiments (internal, 2026)