Can personas evolve in real time to match what users actually want?

PersonaAgent introduces a specific architectural role for the persona concept: a structured system prompt that serves as an evolving intermediary between the agent's memory and its actions. The persona is not static — it evolves continuously by integrating memory-derived insights to guide actions, while action outcomes refine the memory, creating a bidirectional feedback loop.

The architecture has two complementary modules:

Personalized memory module — episodic memory captures detailed, context-rich user interactions; semantic memory generates stable, abstracted user profiles. The persona leverages insights from both memory types to make coherent decisions about how to act.

Personalized action module — the agent's tools and reasoning are tailored to the user. The persona "enforces personalization over the action space and guides action decisions at every step" — it does not merely condition the response but shapes the entire agentic workflow including memory retrieval/update and personalized search/reasoning.

Test-time user preference alignment — the system simulates the latest N interactions, generating responses and comparing them against ground-truth via textual loss feedback. The persona prompt is optimized iteratively through this simulation, ensuring real-time adaptation to the user's current preferences without model retraining. After optimization, learned personas are well-separated in latent space: users with similar interests (e.g., historical/classic films) cluster nearby, while divergent users (e.g., sci-fi/action preferences) show clear separation.

This persona geometry offers a complementary perspective to the Assistant Axis finding. Since How stable is the trained Assistant personality in language models?, PersonaAgent's test-time optimization may work against the Assistant Axis gravitational pull — producing genuine user-specific separation rather than the loose tethering that standard post-training achieves.

A significant limitation: the framework relies on textual feedback for preference alignment, which may overlook implicit or multimodal user signals (emotional or visual cues). This constrains the persona's evolution to what can be expressed and compared in text.

The four-dimension evaluation framework — agentic intelligence, real-world applicability, personal data utilization, and preference alignment — reveals that no prior approach satisfies all four simultaneously. SFT and RLHF achieve general preference alignment but fail individual-level alignment. User-specific fine-tuning achieves personalization but faces computational scaling challenges. Non-parametric approaches have limited data retrieval capabilities.