Agentic Unlearning: When LLM Agent Meets Machine Unlearning

Bin Wang, Fan Wang, Pingping Wang, Jinyu Cong, Yang Yu, Yilong Yin, Zhongyi Han, Benzheng WeiarXiv 2026 · arXiv

TL;DR

This paper introduces agentic unlearning, a new paradigm for completely removing sensitive information from memory-augmented LLM agents. By addressing the "parameter-memory backflow" loop—where external memory re-contaminates updated model weights and vice versa—the proposed Synchronized Backflow Unlearning (SBU) framework successfully purges targeted data across both pathways without destroying the model's general utility.

What Problem Does This Paper Address?

Traditional machine unlearning methods are designed for stateless models, focusing entirely on updating model parameters or modifying inference-time prompts. However, modern LLM agents are rarely stateless; they rely on persistent external memory hierarchies (like vector stores, knowledge graphs, and interaction logs) to maintain long-term context.

If you apply traditional unlearning to these agents, you run into a critical vulnerability the authors call parameter-memory backflow (shown in Figure 1).

Figure 1: Traditional unlearning (left) targets model parameters (θ) only. Agentic unlearning (right) must address both parameters and memory to prevent backflow recontamination (red arrows). Syn- chronized bidirectional forgetting is required.

This creates a dangerous recontamination loop:

  1. Parameter-only deletion: If you scrub a piece of private data from the model's weights, the retrieval mechanism (RAG) might still fetch residual traces from external memory, causing the model to "re-learn" the forgotten data during interaction.
  2. Memory-only deletion: If you clear the external memory store, the model can simply regenerate the sensitive information from its internal weights and rewrite it back into the memory logs.

This dual-retention problem makes isolated unlearning strategies fundamentally insufficient for ensuring compliance with privacy regulations like HIPAA and GDPR.

Methods

To break this loop, the authors propose Synchronized Backflow Unlearning (SBU), a dual-pathway framework that coordinates forgetting across both representations simultaneously.

Figure 2: Overview of the proposed Synchronized Backflow Unlearning (SBU) framework. The framework adopts a dual-pathway design integrating the Memory Unlearning pathway (retrieval-storage) with the Parameter Unlearning pathway (parameters).

The framework consists of three core components:

  • The Memory Pathway (Dependency-Aware Deletion): Instead of naively wiping out records (which might break shared data structures), SBU models memory as a dependency graph. When a delete request arrives, it immediately places the target on an instant blocklist. It then traverses the graph, using reference counting to prune artifacts exclusively derived from the forgotten data while safely preserving shared nodes.
  • The Parameter Pathway (Entropy-Regularized Suppression): Rather than using standard Gradient Ascent (which can cause optimization instability and catastrophic forgetting), SBU uses a KL-to-random alignment scheme. It forces the model's output distribution on forget queries to match a frozen, randomly initialized reference model. This drives the model toward a high-entropy prior, making it maximally uncertain about the forgotten data rather than confidently wrong.
  • Synchronized Dual-Update Protocol: SBU executes memory unlearning first. This ensures that when the parameter pathway updates the weights, it does so using a sanitized retrieval context, preventing the gradients from accidentally re-encoding the target information.

Datasets

The framework was evaluated using II-Medical-8B (a medical LLM built on Qwen3-8B) across three demanding medical QA benchmarks where data privacy is paramount:

  • MedQA: ~10k multiple-choice questions from professional medical board exams.
  • MedMCQA: ~183k multiple-choice questions spanning 2,400 medical topics.
  • MedReason: ~33k question-answer pairs focused on open-ended medical reasoning.

Key Insights and Ideas

  • The Backflow Loop is Real: The paper demonstrates that conventional parameter-efficient baselines (like Sequential LoRA) fail to protect privacy in agents. Even if they seem to forget the data initially, their Membership Inference Attack (MIA) scores remain high because the unchanged memory bank acts as a persistent leak.
  • Uncertainty > Error: Aligning the model to a high-entropy random distribution (making it say "I don't know") is far more stable and effective for knowledge erasure than trying to maximize loss or force incorrect answers, which typically degrades the model's performance on retained data.
  • Order Matters: Synchronizing the updates so that memory is blocked before parameter optimization is crucial. It structurally cuts off the cross-pathway contamination vector during the unlearning process itself.

Key Results

  • Privacy Boost: SBU achieved a Membership Inference Attack (MIA) Score of 0.895 on MedQA (where 1.0 is perfect privacy), representing a 24.8% improvement over standard Sequential LoRA baselines.
  • Utility Preservation: While aggressive baselines like Negative Preference Optimization (NPO) suffered severe generalization collapse (dropping to 41.67% accuracy), SBU maintained >92% test accuracy and 90.5% generalization on retained medical knowledge.
  • Efficiency: SBU scales efficiently as the forget set size increases from 100 to 1,000 instances, maintaining a lower GPU memory footprint than most optimization baselines.

Limitations and Future Work

The authors note that their current dependency-tracking graph mechanism may not fully capture complex, implicit cross-agent information flows within highly interconnected, shared knowledge graphs. As a next step, they plan to explore advanced unlearning protocols specifically tailored for decentralized, multi-agent collaborative environments.

Summary updated

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. Experiments on medical QA benchmarks show that SBU reduces traces of targeted private information across both pathways with limited degradation on retained data.
BibTeX
@misc{wang2026agentic,
  author = {Bin Wang and Fan Wang and Pingping Wang and Jinyu Cong and Yang Yu and Yilong Yin and Zhongyi Han and Benzheng Wei},
  title = {{Agentic Unlearning: When LLM Agent Meets Machine Unlearning}},
  year = {2026},
  eprint = {2602.17692},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2602.17692},
}
Figure