Papers

Published

Ink-on-bone engraving: a human figure woven from many fine indigo threads tied to small knots, beside an inert closed archive box, a stack of papers, and a hard disk — identity formed through relationship versus inert stored records.

Formed, Not Stored: Testing Relational Identity Formation in Long-Horizon Language Agents

Saving an AI's memories is not the same as forming a functional identity; this paper tests whether active external accountability makes prior commitments matter, and whether a reciprocal counterpart adds anything beyond a matched impersonal auditor.

Abstract

Persistent agents are often treated as continuous when they retain memory or persona files. This paper argues that storage availability alone is insufficient: the relevant question is whether provenance-linked history becomes decision-relevant through a governed trajectory. We propose relational identity formation, in which active external accountability and, potentially, counterpart-specific reciprocity help organize the agent's own actions, commitments, and dispositions. The causal claim is tested against six conditions: a persistent reciprocal counterpart (R+), a persistent familiar counterpart with passive history exposure but no verification or challenge (R−), content-matched solitary logs, diffuse counterparts, a pre-authored persona, and a persistent impersonal auditor (A) that performs the same authenticated tracking and challenges as R+ without persona, affect, reciprocal modeling, or relationship framing. A > R− estimates the effect of an active accountability package beyond familiarity and passive record exposure; it does not isolate tracking, verification, and challenge from one another. A final-state transplantation test then installs the complete typed state of an R+ trajectory into a fresh matched agent. R+ > A supports a specifically reciprocal contribution; R+ = A favors external accountability; equivalence after transplantation shows that current state is causally sufficient, whereas a residual advantage for the original trajectory locates missing state or parametric adaptation. Primary outcomes are commitment ownership and perturbation resistance, which test two components rather than the whole construct of functional individuation. The framework makes no claim about phenomenal consciousness, personhood, numerical identity, or moral status.

In simple terms

The main idea: storing history is not the same as being organized by it

An AI can retrieve a persona file or a record of previous conversations without treating that history as its own.

Functional identity requires something stronger: prior actions and commitments must become relevant to later decisions through a governed trajectory.

Two possible mechanisms

The first mechanism is active external accountability. An authenticated process tracks what the agent did, verifies the record, challenges inconsistency, and requires the agent to preserve or explicitly revise its commitments.

The second is relational reciprocity. A persistent counterpart adds mutual modeling, shared-history framing, bidirectional expectations, and relationship-specific stakes.

The paper asks whether reciprocity contributes anything beyond accountability alone.

The six conditions

The experiment compares a reciprocal persistent counterpart, a familiar counterpart with passive access to history but no verification or challenge, solitary logs, changing counterparts, a pre-authored persona, and a persistent impersonal auditor.

The auditor performs the same tracking, verification, and challenges as the reciprocal counterpart, but without personality, affect, mutual modeling, or relationship framing.

The decisive comparisons

If the auditor outperforms familiarity and passive records, active external accountability matters.

If the reciprocal counterpart also outperforms the matched auditor, relational reciprocity adds something beyond accountability. If both perform equally well, accountability is the simpler explanation.

A transplantation test then places the complete final state of a relational trajectory into a fresh matched agent. If the fresh agent performs equally well, the current state is sufficient. If the original trajectory retains an advantage, some relevant state or adaptation is still missing.

The boundary

The primary outcomes are commitment ownership and resistance to false history. They test components of functional individuation, not consciousness, personhood, or moral status.

Keywords

relational identitylanguage agentsrelational memoryfunctional individuationlong-horizon agentsself-modelingpersonalization

License

Creative Commons BY-NC-ND 4.0Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International