Time, Identity and Consciousness in Language Model Agents

Researchers Elija Perrier and Michael Timothy Bennett have introduced a novel framework for evaluating consciousness and identity in AI agents in their paper 'Time, Identity and Consciousness in Language Model Agents,' accepted for the AAAI 2026 Spring Symposium. The core problem they address is that current evaluations of machine consciousness often rely on observing behavior—like an agent's language and tool use—which can be misleading. An agent can say all the right things about itself without those statements being grounded in a persistent, co-instantiated internal state. The authors argue this creates a gap between 'talking like' a stable self and 'being organized like' one.

To bridge this gap, the researchers apply concepts from Stack Theory, specifically the 'temporal gap,' to scaffold and analyze agent trajectories. This method separates when different ingredients of identity occur within an evaluation window from when they are actually co-instantiated at a single decision-making step. By instantiating Stack Theory's Arpeggio and Chord postulates on grounded identity statements, they derive two computable persistence scores from instrumented agent traces. These scores connect to five concrete operational identity metrics, allowing researchers to map different agent scaffolds into an 'identity morphospace' that reveals predictable trade-offs. The result is a practical, conservative toolkit that moves beyond superficial behavioral checks to assess the structural coherence of an agent's self-model over time.