mlsanityData to 14 Oct 2026

Forget to Improve: On-Device LLM-Agent Continual Learning via Budget-Curated Memory

arXiv 2606.25115 ↗First posted 23 Jun 2026cs.LG cs.NIcited by 9

On the map: LLM agents and tool use → Agent threats, guardrails and governance

Abstract (arXiv)

On-device language-model agents improve by accumulating experience in retrieved memory rather than by updating weights. This memory is hard-bounded and exposed: it consumes RAM and energy, reaches peers through a thin uplink, and becomes an attack surface because it is writable by what the agent reads. Existing systems each cover one part of this problem: agentic memories grow without a budget, on-device methods keep entries by success alone, and poisoning is studied mainly as an attack rather than as a memory-governance problem. We propose sys, a single net-value-per-byte score that governs an agent's experience-memory lifecycle. The main idea is to let the budget act as the curator: each entry is scored as value minus harm, per byte, so one ruler decides what to keep, share, and trust. sys makes three decisions: (1) KEEP evicts low-value bytes under the RAM and energy budget; (2) SHARE sends an insight only when its value exceeds its uplink cost; and (3) TRUST gates a peer entry by provenance. On language-model-agent task-drift benchmarks and a real heterogeneous Jetson testbed with two robot-arm nodes and a hub, sys reduces memory by 2.7× and uplink by 2.4×, drives injection success from 0.75 to zero, and raises accuracy on cases corrupted by poison or stale memory. Curating by net value reduces footprint, energy, uplink, and injection success together without reducing accuracy. In this setting, forgetting by net value improves the agent rather than weakening it.

Key facts. Cited by 9 papers in the map (data through 14 Oct 2026). Placed in Agent threats, guardrails and governance, one of 16 directions in LLM agents and tool use. Builds on 50 of its 67 references that are in the map. First posted on arXiv 23 Jun 2026.

Builds on

50 of 67 references.

Closest ancestors · 5

Same direction, newest first

  1. 2026Zombie Agents: Persistent Control of Self-Evolving LLM Agents via Self-Reinforcing Injections17 Feb
  2. 2026Memory Poisoning Attack and Defense on Memory Based LLM-Agents9 Jan
  3. 2025MemoryGraft: Persistent Compromise of LLM Agents via Poisoned Experience Retrieval18 Dec
  4. 2025Memory Injection Attacks on LLM Agents via Query-Only Interaction5 Mar
  5. 2024AgentPoison: Red-teaming LLM Agents via Poisoning Memory or Knowledge Bases17 Jul

Other references · 45

  1. 2024The Llama 3 Herd of Modelscited by 14,715
  2. 2024Qwen2.5 Technical Reportcited by 4,942
  3. 2016Overcoming catastrophic forgetting in neural networkscited by 4,787
  4. 2023Generative Agents: Interactive Simulacra of Human Behaviorcited by 3,568
  5. 2024Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phonecited by 1,917
  6. 2023MemGPT: Towards LLMs as Operating Systemscited by 1,389
  7. 2020ALFWorld: Aligning Text and Embodied Environments for Interactive Learningcited by 1,205
  8. 2025A-MEM: Agentic Memory for LLM Agentscited by 1,063
  9. 2023MemoryBank: Enhancing Large Language Models with Long-Term Memorycited by 556
  10. 2025Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactionscited by 254
  11. 2018BabyAI: A Platform to Study the Sample Efficiency of Grounded Language Learningcited by 244
  12. 2023LLM in a flash: Efficient Large Language Model Inference with Limited Memorycited by 177
33 more
  1. 2024On-Device Language Models: A Comprehensive Reviewcited by 104
  2. 2026Agentic Memory: Learning Unified Long-Term and Short-Term Memory Management for Large Language Model Agentscited by 78
  3. 2023EdgeMoE: Empowering Sparse Large Language Models on Mobile Devicescited by 69
  4. 2025EvoTest: Evolutionary Test-Time Learning for Self-Improving Agentic Systemscited by 37
  5. 2025ShadowCoT: Cognitive Hijacking for Stealthy Reasoning Backdoors in LLMscited by 30
  6. 2025Traceback of Poisoning Attacks to Retrieval-Augmented Generationcited by 28
  7. 2025Apple Intelligence Foundation Language Models: Tech Report 2025cited by 28
  8. 2026Adaptive Memory Admission Control for LLM Agentscited by 23
  9. 2024HiDe-PET: Continual Learning via Hierarchical Decomposition of Parameter-Efficient Tuningcited by 20
  10. 2024Retrieval-Augmented Generation with Estimation of Source Reliabilitycited by 19
  11. 2024Cambricon-LLM: A Chiplet-Based Hybrid Architecture for On-Device Inference of 70B LLMcited by 18
  12. 2026Governing Evolving Memory in LLM Agents: Risks, Mechanisms, and the Stability and Safety Governed Memory (SSGM) Frameworkcited by 18
  13. 2024Prioritized Information Bottleneck Theoretic Framework with Distributed Online Learning for Edge Video Analyticscited by 17
  14. 2026Pretrained Vision-Language-Action Models are Surprisingly Resistant to Forgetting in Continual Learningcited by 17
  15. 2024R-ACP: Real-Time Adaptive Collaborative Perception Leveraging Robust Task-Oriented Communicationscited by 13
  16. 2025Forgetful but Faithful: A Cognitive Memory Architecture and Benchmark for Privacy-Aware Generative Agentscited by 13
  17. 2026Inference-Time Budget Control for LLM Search Agentscited by 13
  18. 2025Cognitive Edge Computing: A Comprehensive Survey on Optimizing Large Models and AI Agents for Pervasive Deploymentcited by 10
  19. 2026When Continual Learning Moves to Memory: A Study of Experience Reuse in LLM Agentscited by 10
  20. 2025Enhancing Vehicular Platooning with Wireless Federated Learning: A Resource-Aware Control Frameworkcited by 9
  21. 2026Agent Memory Below the Prompt: Persistent Q4 KV Cache for Multi-Agent LLM Inference on Edge Devicescited by 8
  22. 2025"X of Information'' Continuum: A Survey on AI-Driven Multi-dimensional Metrics for Next-Generation Networked Systemscited by 7
  23. 2025Shared Spatial Memory Through Predictive Codingcited by 7
  24. 2026RELIEF: Turning Missing Modalities into Training Acceleration for Federated Learning on Heterogeneous IoT Edgecited by 7
  25. 2026Lifecycle-Aware Federated Continual Learning in Mobile Autonomous Systemscited by 7
  26. 2026Reinforcement Learning-Based Energy-Aware Coverage Path Planning for Precision Agriculturecited by 6
  27. 2026Application-Aware Twin-in-the-Loop Planning for Federated Split Learning over Wireless Edge Networkscited by 5
  28. 2026EASE: Federated Multimodal Unlearning via Entanglement-Aware Anchor Closurecited by 5
  29. 2026SCALE: Sensitivity-Aware Federated Unlearning with Information Freshness Optimization for Mobile Edge Computingcited by 5
  30. 2026SPRInG: Continual LLM Personalization via Selective Parametric Adaptation and Retrieval-Interpolated Generationcited by 4
  31. 2026PRISM: Exposing and Resolving Spurious Isolation in Federated Multimodal Continual Learningcited by 3
  32. 2026Learning in the Null Space: Small Singular Values for Continual Learningcited by 1
  33. 2026RAIE: Region-Aware Incremental Preference Editing with LoRA for LLM-based Recommendationcited by 1

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 9.

9 papers.

  1. 2026CrystalMem: Elastic Memory for Self-Evolving LLM Agents via Knowledge Crystallizationcited by 2
  2. 2026From Cognitive Architectures to Language Agents: A Mechanism-Level Review of Lineage, Convergence, and Migration Gapscited by 1
  3. 2026ERRAND: Budgeted Maintenance of Agent Memorycited by 1
  4. 2026When Unlearning Fails: Reliable Data Deletion under Post-Training in Agent Networkscited by 0
  5. 2026Continual Learning in Transitioncited by 0
  6. 2026Not All Experience Belongs in the Weights: Component Routing for Self-Improving GUI Agentscited by 0
  7. 2026Causal Memory Policy: Making Memory Utility Identifiable by Intervening on Retrievalcited by 0
  8. 2026Curating Always-Loaded Context for LLM Agents: A Capacitated Assortment Model with Censored Feedbackcited by 0
  9. 2026Recurrent Harness for Agentic Continual Learningcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026Bad Memory: Evaluating Prompt Injection Risks from Memory in Agentic Systems
  2. 2026From Attack Success to Attack Severity: Counterfactual Memory Attacks on LLM Agents
  3. 2025How Memory Management Impacts LLM Agents: An Empirical Study of Experience-Following Behavior
  4. 2026Memory as a Controlled Process: Learned Adaptive Memory Management for LLM Agents
  5. 2026From Untrusted Input to Trusted Memory: A Systematic Study of Memory Poisoning Attacks in LLM Agents
  6. 2026Memory as Middleware for Self-Improving AI Agents
  7. 2026SMSR: Certified Defence Against Runtime Memory Poisoning in Persistent LLM Agent Systems
  8. 2026Compiled Memory: Not More Information, but More Precise Instructions for Language Agents