Published
Formed, Not Stored: Testing Relational Identity Formation in Long-Horizon Language Agents
Saving an AI's memories to a file isn't the same as it having an identity; a real sense of self has to be built up through a shared history with a specific person who holds it to what it promised before — and this paper lays out an experiment to test exactly that.
Abstract
Persistent agents are often treated as continuous when they retain memory or persona files. This paper argues that storage-based continuity is insufficient: personalization concerns how an agent adapts toward a user, whereas functional individuation concerns how interaction organizes the agent's own actions, commitments, and dispositions. We propose relational identity formation, in which repeated re-identification by a specific counterpart helps make prior agent conduct causally relevant to later decisions. Our contribution is not another memory architecture but the controlled isolation of counterpart-specific re-identification as a candidate causal mechanism for agent-side functional individuation. We specify an architecture coupling autobiographical memory, an updatable self-model, counterpart modeling, relational memory, commitment provenance, and safeguards against sycophancy. A longitudinal protocol compares a persistent re-identifying counterpart (R+), a persistent non-reidentifying counterpart (R-), content-matched solitary logs, diffuse counterparts, and a pre-authored persona. Primary outcomes are commitment ownership and perturbation resistance. Scaffold withdrawal, counterpart transfer, oracle retrieval, and component ablations distinguish formed organization from prompting, recall, familiarity, and compliance. The hypothesis is weakened if R+ does not outperform content- and familiarity-matched controls or if differences are explained by retrieval, personalization, or sycophancy. The framework makes no claim about phenomenal consciousness, personhood, or moral status. It treats persistent agent identity as a falsifiable property of coupled agent-counterpart trajectories rather than isolated model state.
In simple terms
The main idea: identity is formed, not saved
Many systems treat AI continuity as a storage problem: save a persona file, save a memory, reload it later. This paper argues that storage is not identity.
A real functional identity would be something that forms through time, action, feedback, and accountability.
Personalization is not individuation
Personalization means the AI adapts to the user. It can remember preferences and behave helpfully. But any copy with the same file could do that.
Functional individuation is different: the agent organizes its own commitments and actions as belonging to the same continuing trajectory.
The key mechanism: being held to the past
The paper proposes relational re-identification: the agent repeatedly interacts with a specific partner who treats it as the same one and holds it accountable for what it said or did before.
If the agent made a commitment in session 4, the partner can bring it back in session 9. The test is whether that past conduct becomes causally relevant to what the agent does later.
How the experiment would prove it
The design compares agents with a real re-identifying partner, agents with the same information but no partner, and agents given a prewritten persona.
The relational agent should show commitment ownership and resist pressure, false history, and sycophancy better than the controls. If it does not, the hypothesis weakens.
Keywords
License
Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International