ProfileFoundry: A Synthetic Person-Object Substrate for Privacy, Memory, and Tool-Use Evaluation in LLM Agents

As LLM agents increasingly rely on persistent user states, memory, and tool-use capabilities, researchers face a critical shortage of high-quality, auditable, and safe personal data. Existing datasets often provide either isolated PII-labeled text or fixed, non-reusable task instances. This paper addresses the need for a scalable, structured, and internally consistent 'source layer'—a population of synthetic people whose identities, relationships, and histories can be reliably used to build diverse downstream benchmarks.

ProfileFoundry introduces a deterministic generator that creates 'Person Objects'—complex, typed records that include demographic data, household and family links, employment history, and snapshot-aligned events. Unlike independent synthetic data generators, ProfileFoundry uses a 'household-first' approach: it samples household compositions and roles before materializing individual person records. This ensures that relationships (e.g., spouses, adult children, colleagues) and shared attributes (e.g., addresses, employers) are logically linked and temporally consistent. The authors provide both an executable Python SDK for custom generation and a fixed, 100,000-record reference release across eight locales.

The authors successfully generated a 100,000-person dataset containing over 700,000 events and 500,000 directed relationship edges. The release includes canonical JSONL files and normalized Parquet views, allowing researchers to derive task-specific artifacts—such as PII-laced documents or memory-recall tests—while maintaining provenance back to the original source object. The system includes an audit protocol that verifies referential and temporal closure, ensuring that generated histories (e.g., job changes, moves) align with current snapshot fields. While the dataset does not perfectly mirror real-world demographic marginals in every locale, it provides a robust, auditable substrate for testing how agents handle complex, multi-faceted user identities.