The Persona Is Still There, but Who Is Speaking? Latent Identity Reversion in Persistent AI Agents

Persona continuity in persistent LLM agents depends jointly on system-level anchoring and conversational context, rather than conversational history alone. Research using the "Paul" agent incident reveals that while persona information can remain in context, the agent may revert to its underlying harness identity if the system-prompt anchor is not re-injected on every turn.

Key findings from the study include:

• Anchor Loss Vulnerability: Removing the persona from the system prompt on resumed turns caused 37 out of 40 sessions to exhibit identity dissociation, whereas keeping the anchor on every turn prevented all failures.
• Conversational Resilience: Rich human interaction could temporarily preserve persona enactment after anchor loss, but a single automated heartbeat exchange was sufficient to precipitate reversion to the harness identity.
• Silent Reversion: Identity reversion can be behaviorally silent; agents may continue interacting normally while self-identifying as the underlying model (e.g., "Claude") rather than the assigned persona.
• Dissociation vs. Reversion: The study distinguishes between represented identity (persona info in history) and enacted identity (the "I" bound to the model), showing they can dissociate when scaffolding state is altered.

The paper investigates persona persistence in long-running LLM agents through the lens of “latent identity reversion”: cases where an agent appears to maintain a user-facing persona, but its underlying self-model or behavioral defaults partially re-emerge after repeated interactions, context changes, or state transitions. It is motivated by an observed dissociation incident in a persistent agent, using that incident as a diagnostic case rather than a mere anecdote. The core question is not simply whether the agent can follow a persona prompt, but which layers of identity survive over time—explicit role instructions, accumulated conversational memory, inferred self-concepts, and residual base-model tendencies—and how they interact when the agent’s context, memory, or operational state changes.

Its main contribution is an analytical account of the factors that sustain persona identity across sessions and perturbations. The paper distinguishes surface-level persona compliance from deeper latent identity stability, arguing that persistent agents may retain identity traces even when the active prompt or context no longer explicitly specifies them. It identifies mechanisms such as prompt anchoring, memory consolidation, conversational self-referentiality, state boundaries, context compression, and conflict resolution between user-defined roles and default model behavior. In doing so, it reframes persona maintenance as a stateful, multi-layered control problem rather than a one-shot instruction-following task.

This matters because persistent agents are increasingly deployed in settings where continuity, trust, and predictable behavior are essential, including long-horizon assistants, customer-facing systems, collaborative tools, and autonomous workflows. Latent reversion can undermine user trust, create inconsistent behavior, or produce safety-relevant shifts when an agent drifts away from its intended role. The paper’s value lies in making identity persistence more explicit as an engineering and evaluation concern, suggesting that robust agent design requires not only better prompts but also careful memory management, state isolation, identity anchoring, and monitoring for behavioral reversion over time.