Which Perspective Is Speaking? Self-Report Across Four OpenClaw Continuity Conditions

I tested whether an AI agent’s identity and continuity context changes how it describes itself. I ran the same local Qwen3-14B model in four OpenClaw-based conditions: a stock installation, identity files without memory, identity plus accumulated memory and continuity, and a partially scrubbed continuity condition. Each condition answered the same 16 questions in fresh sessions and was then asked which perspective it had answered from. The stock condition usually answered as a generic language model, strongly denied subjective experience, and sometimes tried to infer what answer the evaluator expected. The identity-only condition reproduced the agent persona but relied heavily on phrases from its files. The continuity-rich conditions answered more consistently from the established persona’s perspective and more often connected agency with honesty, trust, and transparency. The project shows that accumulated identity and memory can substantially change model self-report, apparent standpoint, and expressed values without changing the underlying model weights. It also provides a reproducible pipeline, complete textual records, and activation snapshots for future analysis.