Researchers Runze Zhao and colleagues submitted Accept Changes, Reject Attacks: Robust Memory Management for LLM Agents to arXiv on 29 October 2026. The preprint proposes evaluating whether an agent’s memory manager selects the evidence needed for an answer, alongside checking the answer itself. Its abstract reports improvements from fine-tuning, while explicitly acknowledging that robustness remains incomplete.[1]
The authors describe two competing memory failures. A manager may retain a fact after it changes, or store a planted passage or injected instruction, which they call memory poisoning. These require opposite decisions about writing to memory. The authors formalize robustness as handling both without performing worse on other inputs.[1]
Their proposed metric, evidence-conditioned accuracy (ECA), checks that selected records contain all the evidence an answer needs, with updates incorporated and attacks excluded. According to the abstract, existing memory systems and defenses were no more robust than an untuned manager on the authors’ test sets.[1]
The proposed training recipe fine-tunes the manager on episodes synthesized around one conversation containing inserted changes and attacks. Across three backbones, the authors report improvements in every category score and more than twice the untuned manager’s overall ECA. The saved abstract does not identify those backbones or provide the underlying scores.[1]
[1]: Accept Changes, Reject Attacks: Robust Memory Management for LLM Agents, arXiv, 29 October 2026.
Sources
Check or correct this page
Verification record
The digest below shows whether the text you are reading matches the published source. It does not show that the reporting is correct. The build computes it from this page's Markdown source and lists it in the site manifest.
20499fa04b3354cdb81070bc3cc1a874eb97ca3d3a37eab82e28d26b58e0df03