In evaluating AI companions, not only single-turn fluency matters, but also sustained emotional continuity, which includes remembering the companion's identity, user preferences, and the feeling of their relationship. We introduce ZifaMem, a structured memory system that organizes dialogues into session summaries, episodic memories, and a consolidated user model.
Against a deployment-honest comparator that supplies the full raw dialogue history, under a fixed LLM-as-a-judge protocol, structured memory raised pooled four-backbone emotional-intelligence scores by 11.4% (95% CI 6.3% to 17.1%). Persona grounding improved across all four backbones (Claude +42% relative). Multi-turn affect context achieved a +39% net preference over a single-turn snapshot (exploratory), while an additional emotion state machine yielded no measurable gain on any of five endpoints.
Under an identical preregistered protocol, three memory systems (ZifaMem, Mem0, and filtered verbatim retrieval) each significantly outperformed raw-history deployment, with ZifaMem and Mem0 being statistically equivalent within +/-5 points on the preregistered primary preference endpoint. The ZifaMem SDK, CLI, and portable Agent Skills are open-sourced at GitHub.
Blogger's Review: ZifaMem's structured memory system adds a new dimension to emotional interactions in AI companions, significantly enhancing user experience. By effectively integrating dialogue history, it not only improves understanding of user preferences but also ensures emotional continuity, making it a noteworthy advancement in practical applications.