• Computer Science > Machine Learning [Submitted on 6 Feb 2026] Title:Agentic Unlearning: When LLM Agent Meets Machine Unlearning View PDF HTML (experimental)Abstract:In this paper, we introduce \textbf{agentic unlearning} which removes specified information from both model parameters and persistent memory in agents with closed-loop interaction. • Existing unlearning methods target parameters alone, leaving two critical gaps: (i) parameter-memory backflow, where retrieval reactivates parametric remnants or memory artifacts reintroduce sensitive content, and (ii) the absence of a unified strategy that covers both parameter and memory pathways. • We present Synchronized Backflow Unlearning (SBU), a framework that unlearns jointly across parameter and memory pathways. • The memory pathway performs dependency closure-based unlearning that prunes isolated entities while logically invalidating shared artifacts. • The parameter pathway employs stochastic reference alignment to guide model outputs toward a high-entropy prior. • These pathways are integrated via a synchronized dual-update protocol, forming a closed-loop mechanism where memory unlearning and parametric suppression reinforce each other to prevent cross-pathway recontamination.

Article Summaries:

  • Agentic Unlearning: Removing Sensitive Data from Both Parameters and Memory in LLM Agents

Researchers have introduced agentic unlearning, a method that deletes specified information from both the internal weights and the persistent memory of large‑language‑model agents. Existing techniques target only model parameters, leaving “back‑flow” risks where memory retrieval or re‑injected artifacts can re‑introduce sensitive content. The new framework, Synchronized Backflow Unlearning (SBU), jointly prunes memory via dependency‑closure and suppresses parameters through stochastic reference alignment, coordinated by a synchronized dual‑update protocol. Experiments on medical question‑answering benchmarks show that SBU effectively removes private data from both pathways while preserving overall performance.

Sources: