Research
The Price of Safety: Benign-Case Utility and Token Overhead of Memory-Poisoning Defenses in LLM Agents
Publication date not yet evaluated · Added arxiv.org
Page published
Coverage timeline
Single-source research — one report is available.
Why it matters
Memory-poisoning defenses run on all agent traffic while attacks are rare, so quantifying their cost on benign interactions helps defenders decide whether a defense degrades ordinary agent operation more than it protects it.
A research paper by Pritom Bhowmik measures the benign-case utility and token overhead of four memory-poisoning defenses for LLM agents (input sanitization, provenance checking, LLM-based anomaly detection, and retrieval-time reranking) on entirely benign LoCoMo traffic. Write-time defenses show no resolvable utility cost, while the read-time reranker lowers core accuracy by 4.4 points and quarantines legitimate memories on 33.6% of adjudicated items at 2.7% token overhead. Code is available at github.com/pritom02bh/memdefense.