Federated Forgetting in Agentic Workflows: GDPR Compliance Experiments with Synthetic User Logs

Zichao Li, Zong Ke

Keywords: federated learning, agentic workflows, GDPR compliance, machine unlearning, differential privacy, temporal modeling, synthetic benchmarks

TL;DR: This paper develops a federated forgetting framework for agentic workflows that achieves GDPR compliance through temporal influence quantification, memory-buffered scrubbing, and DP verification, with 92% forgetting completeness.

Abstract: This paper introduces a novel framework for GDPR-compliant federated forgetting in agentic workflows, addressing three key challenges: (1) temporal influence quantification through windowed gradient analysis, (2) privacy-preserving scrubbing with memory buffers, and (3) differential privacy verification. Our method achieves 92\% forgetting completeness on WebArena (13.6\% improvement over baselines) while maintaining 91\% accuracy on retained knowledge and 98\% GDPR compliance. Experiments across six benchmarks demonstrate practical deployment viability with 136ms/request overhead. The solution bridges critical gaps in adaptive workflow management, regulatory compliance, and privacy-preserving benchmarking for federated agentic systems.

Submission Number: 21