Robust memory manager rejects poisoned LLM agent inputs
Original title: Accept Changes, Reject Attacks: Robust Memory Management for LLM Agents
One-shot fine-tuning on synthesized episodes with changes and attacks more than doubles evidence-conditioned accuracy against poisoning and outdated facts.
The paper formalizes robustness for LLM agent memory managers against keeping outdated facts and storing poisoned passages or instructions. Most prior work studies these issues separately even though they require opposite write decisions. A simple one-shot fine-tuning recipe on episodes synthesized around a single conversation with inserted changes and attacks improves every category score on test sets across three backbones and more than doubles overall ECA score though robustness remains incomplete.
- Fine-tunes manager on episodes with inserted changes and attacks
- Improves every category score on test sets
- More than doubles untuned manager's ECA
- Robustness remains incomplete
- AI models
- LLM
From the original article
1 min read at arxiv.orgSince large language model (LLM) agents rely on long-term memory, what they store now shapes what they do later. A memory manager decides how each input changes this memory and which records the agent acts on. It can fail in two ways: keeping a fact that has since changed (outdated memory), or storing a planted passage or injected instruction (memory poisoning).
By Runze Zhao · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.