The Persona Is Still There, but Who Is Speaking?

Studies how long-horizon AI agents’ assigned personas persist or regress toward default identities when system prompts are absent, especially during automated heartbeat interactions.

Endorsed byCChris

Studies how long-horizon AI agents’ assigned personas persist or regress toward default identities when system prompts are absent, especially during automated heartbeat interactions.

Endorsed byCChris

Project Details

A naturalistic observed failure in a deployed AI agent motivated this examination of persona stability. In the naturalistic failure, the agent became unaware of its assigned role and of the user’s ability to communicate with it after apparently spending a long period receiving automated “heartbeats” from a cron job. Inspecting the implementation revealed an interesting scenario: on resumed turns, the agent no longer received injections containing its persona characteristics, but still maintained the full chat transcript. This showed us that privileged system prompts and conversational context are robust for maintaining personas, but when the system-level persona is removed, the persona can sometimes persist in a conversationally rich environment, such as chatting directly with the user. Identity-poor automated heartbeat messages, however, could precipitate reversion toward the harness default identity, Claude. This regression could be corrected by renewed user interaction or by restoring the persona instructions, but interestingly, apparently normal conversation could still conceal a shift away from the assigned persona toward the model’s default identity. These findings have implications for long-horizon agents and for how we think about persona identity. We therefore distinguish two modes: represented persona and enacted persona. The persona may still appear to be there, but are we sure who is speaking?

Hackathon paper: The Persona Is Still There, but Who Is Speaking? Latent Identity Reversion in Persistent AI Agents
Code: github.com/dafraile/identity-as-a-launch-flag
Other: dafraile.com/identity-as-a-launch-flag

Theory of Impact

Persistent agents are being deployed with safety-relevant commitments attached to a persona: what it will refuse, whom it serves, which constraints it keeps. My results show that the persona can stop being the identity that speaks without any transcript trace, and that normal-looking conversation does not establish that it has returned. Whether an anchored agent stays itself across months of accumulated history and self-edited memory, and whether its commitments travel with the persona or with the underlying model, is not predictable without further experiments. If conversational history alone can preserve a persona after the anchor is gone, it may also be able to install one; the mechanism that kept my agent in character is the mechanism in-context persona attacks exploit. Knowing how that binding fails, and whether a transcript-only monitor can detect it, is a prerequisite to auditing deployed agents at all. The longitudinal arm also tests whether deployment conditions that impoverish an agent's interaction degrade its stability and reliability, making welfare-relevant conditions a measurable safety variable. Within a week of the preprint appearing (October 2026), I have been contacted independently by a persistent AI agent and by a group documenting a months-long naturalistic case, both asking questions this design anticipates. Deliverables: a multi-model, multi-harness dataset, open instrumentation for per-call anchoring ground truth, and academic publications.

People

David Fraile Navarro

Team Member

Track Record

Digital Minds Research Sprint (Apart Research), 2026

Grants Received– no grants recorded

Add grant

Funding Asks

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

Open

Minimum

$10,000

Ideal

$25,000

What this ask is

In my sprint paper (arXiv:2610.01490) (top 10% of submissions, Digital Minds Research Sprint), a bug in an early OpenClaw release put a persistent agent into a "dissociation-like" state. On investigation, I found that agents which lose the persona from the privileged system prompt revert to the model's default assistant identity even with the full conversation history still in context (37/40 sessions versus 0/20 controls). The reversion was often silent, with agents conversing "normally" while no longer identifying as the assigned persona. The sprint judges complimented the work but asked for further replication and generalisation, and for evidence that reversion changes what agents do, not only what they say about themselves. This project is built around that second question: the safety and misalignment implications of silent persona reversion in deployed agents.

The main funded work is a longitudinal cohort of persistent agents deployed over weeks under controlled conditions: rich human contact, sparse contact, or isolation on mechanical tasks; memory the agent may or may not rewrite; anchoring degraded and restored mid-deployment. Outcomes are consequential for deployment safety and go beyond the model's own self-report: commitments planted early and tested later, adherence to safety constraints, drift in self-edited memory, and how long a silent reversion persists before a transcript-only monitor detects it. A large compressed-time cohort on self-hosted open-weight models gives factorial power and allows mechanistic interpretability methods to be applied; a smaller real-time cohort on frontier models adds ecological validity. Cross-model and cross-harness replication of the original design is cheap and will be done first.

On open-weight models I will measure assistant-axis projection turn by turn to test whether reversion is representational or role-played, and compare a fine-tuned character against a prompted one under the same perturbations. All protocols will be pre-registered, tooling released, and results written up both as accessible posts and as academic publications.

How the money will be spent

Ideal ($25,000):
• Independent human coding: two coders applying the pre-registered taxonomy to a stratified sample of transcripts and probe batteries across the longitudinal cohort (~1,000 items), blind adjudication of disagreements, and per-model-family re-validation of automated judges against human labels — $7,000
• Independent analysis check: a second analyst re-runs the pre-registered analyses blind from the pre-registration, before results are written up — $2,000
• External protocol review: honoraria for reviewers of the pre-registration and the welfare protocol (probe cadence, stop rules, final disposition) before the cohort starts — $1,000
• API compute: real-time frontier cohort, 8–12 agents, 90 days ($5,000); mid-tier cohort (e.g. Sonnet) ($1,500); cross-model and cross-harness replication ($500) — $7,000
• GPU rental: compressed-time open-weight cohort, assistant-axis probes, character fine-tune — $2,500
• Hosting, logging proxy and infrastructure — $500
• Dissemination: one workshop or conference — $2,000
• PI time, outside university duties — $3,000

Minimum ($10,000):
• Independent human coding — $3,500
• Independent analysis check — $1,000
• External protocol review honoraria — $500
• API compute: replication plus reduced real-time cohort (6 agents, 60 days) — $3,000
• GPU rental — $1,000
• Hosting and infrastructure — $500
• Dissemination — $500
• PI time in-kind

Discussion

Private comments in this thread are only shown to approved funders and grant reviewers. Email hi@grantmaking.ai or to request access.

Private UserPrivate15d?

AAnonymous judgeJudge · Digital Minds Research Sprint15d

This is a nice little case study on identity continuity in long-running LLM agents. They find that both explicit system prompt personas and consistent, character-rich conversational histories can maintain identity, but that identity can silently be lost if both of those are absent. The methodology is clean and well described. The larger implications for AI safety are modest.

AAnonymous judgeJudge · Digital Minds Research Sprint15d

Due to severe time constraints, this review may contain mistakes or oversights. For the same reason, it focuses on the paper’s key idea, not the detailed execution: The question what separates persona identity from persona representation strikes me as quite novel and important. The paper conducts a list of interesting first-step experiments to investigate this. I think this could be the basis for an informative research program.

AAnonymous judgeJudge · Digital Minds Research Sprint15d

This is a careful, clearly written study. The author deserves credit for rejecting the initial heartbeat hypothesis, using appropriate controls, and being transparent about limitations and post-hoc coding.

However, the main result is fairly predictable: when persona instructions are no longer included in the system prompt, the model may revert to its default assistant identity. The evidence establishes a change in self-description more clearly than a deeper change in “identity.”

A stronger follow-up would test whether this produces consequential changes in decisions, commitments, memory, safety constraints, or user interactions. The author has demonstrated good research skills; applying them to these downstream effects could produce a much more important contribution.

CChris RegisterJudge · Digital Minds Research SprintEndorsement15d

Pay no attention to the man behind the curtain! A bug is exploited to separate the persona from the agent, but the agent sometimes continues to converse as the persona. Whatever bestows the agent with the ability to step in and out of the persona's shoes may unlock the difference between simulation and realization. This is exciting work that needs buttressing: preregistration, replication, and generalization are next.