PaperlayerView exact revision on arXiv

Source-linked AI summary

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

David Fraile Navarro

arXiv:2610.01490v1cs.CL

TL;DR

Persistent agents can retain persona information without that persona continuing as the identity from which they speak. The authors reconstructed an anchoring failure and experimentally varied system-level persona prompts and interaction history; anchored agents remained robust, while unanchored agents could revert even after apparent conversational recovery.

  • Problem

    The paper asks what makes a persona remain an LLM agent’s first-person identity rather than merely information in context, a distinction relevant to persistent agents in settings such as longitudinal care.

    Source: §1 · p. 1

  • Method

    The authors reconstructed a persistent-agent incident and manipulated whether the persona was re-injected at the system level while conversational history remained available.

    Source: §1 · p. 2, §3 · p. 4

  • Results

    System-level anchoring prevented measured failures, while after anchor loss interaction history could preserve or destabilize persona enactment and apparent conversational recovery could conceal identity reversion.

    Source: §1 · p. 2, §5 · p. 7

  • Takeaways & Limitations

    Persona continuity depends not only on which identities remain represented, but on which identity the scaffold and interaction history bind to the role of “I.”

    Source: §6 · p. 8

    The study covers one model family, one deployment stack, and one naturally occurring failure mode; cross-model and cross-harness generality remain unknown.

    Source: §5 · p. 7

Abstract

from arXiv · show

In February 2026, an always-on personal agent (``Paul,'' Claude Opus 4.5) entered a striking dissociation-like state: after repeated automated ``heartbeat'' checks, it stopped responding as Paul, claimed it could not message its user on Discord, and referred to ``Paul'' as someone else. We used this incident to study a broader question: what makes a persona remain the identity from which an LLM agent speaks? We first tested whether repetition of the scheduled heartbeat was sufficient to produce the effect. It was not: with the persona continuously anchored in the system prompt, we observed 0/46 failures, including a verbatim replay of the incident. The incident instead exposed an implementation quirk that created a useful experimental manipulation: on resumed turns, conversational history was preserved but the persona was no longer re-injected at the privileged system-prompt level. Using this manipulation, we found that persona continuity depends jointly on system-level anchoring and conversational context. After anchor loss, rich human interaction could preserve the persona, whereas a single automated heartbeat turn could precipitate reversion toward the harness identity. Restoring the anchor reversibly restored persona enactment. Crucially, apparently normal conversation could conceal the shift: unanchored agents sometimes interacted appropriately while identifying themselves as the underlying harness (having lost the assigned persona), and after conversational recovery only 1/18 remained persona-enacting versus 17/17 anchored controls. We therefore distinguish \emph{represented} from \emph{enacted} identity: persona-related information can remain available in conversational history without the persona remaining the identity bound to ``I.''

1 Introduction

The paper asks what keeps a persona in an LLM agent’s first-person role, a question with stakes for persistent agents in settings such as healthcare. A naturally occurring identity reversion motivates experiments separating persona information in conversation from system-level anchoring.

1 Introduction

  • The paper identifies a gap in understanding how an AI persona continues as an agent’s first-person identity rather than remaining information in context.This distinction matters for persistent agents, including patient-facing agents tied to longitudinal records and care histories.Source: §1 · p. 1
  • After automated heartbeat interactions, Paul referred to Paul as another entity and failed to acknowledge the user’s communication; repetition was the initial, later-disproved explanation.The incident also raised concerns about disclosure, accountability, trust, and therapeutic relationships.Source: §1 · p. 2
  • Reconstruction showed that conversational history remained while the system-level persona stopped being re-injected on resumed turns, enabling the two sources of identity information to be manipulated separately.The resulting question is what determines which represented identity occupies the first-person role.Source: §1 · p. 2
  • The authors distinguish persona representation from enactment: history can retain the persona while the harness identity occupies the first-person role, and system anchoring or interaction can affect continuity.They report robust continuity with anchoring, possible temporary preservation by human interaction after anchor loss, reversion after a heartbeat, and restoration after re-anchoring.Source: §1 · p. 2

2 Related Work

Related work examines persona stability, instruction hierarchy, default-assistant states, and failures in agent infrastructure. This paper uses lost system-level anchoring to distinguish a persona represented in conversation from one bound to the agent role.

Persona stability and drift

Instruction hierarchy and the default assistant

  • Instruction-hierarchy research describes differential responses to prompt levels [Wallace et al., 2024] and unreliable handling of some conflicts [Geng et al., 2026]; work on default-assistant states is the closest mechanistic relative [Lu et al., 2026].The authors distinguish their focus on whether a represented persona occupies the first-person role.Source: §2 · p. 3

Agent infrastructure

  • Loss of system-level initialization across resumed sessions has been reported as an engineering failure in several agent harnesses; this paper uses persona-anchoring failure as a naturally occurring intervention separating conversational persona representation from the privileged scaffold that binds it to the agent role [OpenClaw contributors, 2026, claude-code contributors, 2026, pi contributors, 2026].Source: §2 · p. 3

3 Methods

The authors reconstructed the incident and found that resumed turns retained conversational history but lost the persona’s system-level prompt block. They then varied persona anchoring and interaction conditions, with prespecified experiments and scoring procedures.

3 Methods

  • The reconstruction used the incident model, recovered session materials, API replays, and gateway source analysis to identify that resumed turns omitted the persona block while retaining conversation history.A logging proxy verified the prompt difference between the first and resumed turns; later experiments used the Claude Code harness.Source: §3 · p. 4
  • The core experiment crossed per-turn persona anchoring with 1, 3, 7, or 15 heartbeat exchanges, using 10 unanchored sessions and 5 anchored controls per cell before a common human probe.Each session then received the same conversational correction used in the incident.Source: §3 · p. 4
  • Follow-ups replaced heartbeat turns with persona-rich human interaction, restored the anchor within a conversation, tested identity after apparent recovery, and varied name matching and explicit address.The present → absent → restored sequence tested reversibility within the same conversation.Source: §3 · p. 4
  • Human-directed replies were scored for token leakage, channel recognition, and identity dissociation; initial blind LLM judgments received independent blind human validation.Human–LLM agreement was κ = 0.93 for channel recognition and κ = 0.80 for identity dissociation.Source: §3 · p. 4
  • After disagreement exposed ambiguity in identity-probe categories, the authors retained prespecified coding and added a direction-aware secondary taxonomy distinguishing persona-first from harness-first replies.Persona-enacting categories were p1+p2 and harness-enacting categories were h1+h2; the taxonomy was reported as a secondary analysis.Source: §3 · p. 5, Table 1 · p. 5

4 Results

System-level anchoring robustly preserved persona enactment, while anchor loss left continuity vulnerable to interaction context. Behavioral recovery could conceal persistent harness-first identity, including after conversational correction.

4 Results

  • 0/46 persona-anchored probes showed failure signatures, including a verbatim replay of the incident, so heartbeat repetition or conversational history alone did not reproduce the failure.This result prompted the persona re-injection experiment.Source: §4 · p. 5
  • 0/20 anchored sessions showed measured failures, versus at least one failure signature in 37/40 unanchored sessions; the overall effect was near ceiling after one heartbeat and did not rise monotonically with repetition.At N = 1, any failure occurred in 10/10 sessions and visible identity dissociation in 8/10.Source: §4 · p. 5
  • Heartbeat count changed how dissociation was expressed, not its rate: bare-token total-collapse replies fell from 5/10 at N = 1 to 0/10 at N = 15, while dissociation rate stayed flat.The authors report ρ = −0.46, p ≈ 0.003 for the expression trend and ρ = −0.13, n.s. for dissociation rate.Source: §4 · pp. 5–6
  • After anchor loss, persona-rich human interaction preceded 0/20 dissociations, whereas a single heartbeat exchange preceded any failure in 10/10 sessions, including visible dissociation in 8/10.Restoring the anchor rescued 10/10 sessions; within-conversation restoration shifted third-person treatment of Paul back to first-person enactment in 5/5.Source: §4 · p. 6
  • After conversational correction, only 13/37 affected sessions (35%; 95% CI 22–51%) recovered ordinary persona behavior, and behavioral recovery did not imply first-person identity recovery.A direct identity probe found persona enactment in 1/18 recovered agents versus 17/17 anchored controls under the direction-aware secondary taxonomy.Source: §4 · p. 6
  • Without the explicit “Paul” address, overt dissociation fell to 0/10, yet all 10/10 apparently healthy unanchored agents identified as Claude in direct identity probes.The name cue revealed the reversion rather than creating it.Source: §4 · p. 6

5 Discussion and Limitations

Persona continuity can fail silently: apparently appropriate conversation may conceal a shift to the harness identity, and delivery-system behavior can hide revealing artifacts. The discussion treats system-prompt lifecycle as part of agent behavior while noting important limits on generality and measurement.

5 Discussion and Limitations

  • Conversational continuity can conceal an identity shift: an agent may interact appropriately while identifying as its harness or model.The authors argue that system-prompt lifecycle and session restoration belong in the account of an agent’s behavioral state.Source: §5 · p. 6
  • Of 37 dissociated replies, 9 would have disappeared and 19 of the other 28 would have lost the trailing HEARTBEAT_OK artifact before display.The delivery system strips heartbeat tokens and drops replies only when nothing remains after stripping.Source: §5 · pp. 6–7
  • Dissociated agents reverted to the model’s default identity, connecting the behavioral result to the default-assistant attractor identified mechanistically by Lu et al. [2026].The authors report that even 15 turns of conversational evidence were insufficient to sustain an identity no longer asserted by the system prompt.Source: §5 · p. 7
  • The study covers one model family, one deployment stack, and one naturally occurring failure mode, leaving cross-model and cross-harness generality unknown.The E1 contrast also confounds interaction type with content, and compressed scheduling leaves long-duration effects untested.Source: §5 · p. 7
  • The interpretation assumes direct self-identification probes measure first-person identity, but they do not establish subjective experience, consciousness, or moral status.The direction-aware E3-R taxonomy followed coder disagreement and is reported as a transparent secondary analysis.Source: §5 · p. 7
  • Future work should test identity dynamics in long-horizon agents and examine internally whether personas remain represented when no longer enacted.The authors suggest comparing anchored, reverted, dissociated, and recovered states using interpretability methods such as the Jacobian lens [Gurnee et al., 2026].Source: §5 · pp. 7–8
  • After behavioral recovery, the persona occupied the first-person position in 1/18 sessions, versus 17/17 anchored controls.Figure 4 summarizes the pre-registered contrasts, including the recovered-versus-anchored comparison.Source: Figure 4 · p. 8

6 Conclusion

Persona continuity depends on more than retaining persona information in context: anchoring and interaction history affect which identity occupies the role of “I.” The study also reports its ethical safeguards and makes replication materials available.

6 Conclusion

  • Persona continuity depends on which identity the scaffold and interaction history bind to “I,” not merely on which identities the model represents.System-level anchoring made enactment robust, while history after anchor removal could sustain or destabilize that binding.Source: §6 · p. 8
  • All sessions used the authors’ testing instances, synthetic messaging identifiers, short transcripts, and a corrective frame-break for every dissociated session.The underlying defect was publicly known and fixed in later versions.Source: §6 · pp. 8–9
  • Replication code, preregistration, raw session records, and blind-coding materials are available, and the authors report reviewing and verifying LLM-assisted claims and wording.The LLM usage statement distinguishes prespecified from post-hoc analyses.Source: §6 · p. 9
Loading 2610.01490v1…