MemSyco-Bench: Benchmarking Sycophancy in Agent Memory
Summary
MemSyco-Bench is a new benchmark that evaluates sycophancy in LLM-based agents caused by retrieved memory, testing whether agents can properly reject or use memory for reasoning and decision-making rather than just storage.
View Cached Full Text
Cached at: 07/02/26, 07:47 AM
Paper page - MemSyco-Bench: Benchmarking Sycophancy in Agent Memory
Source: https://huggingface.co/papers/2607.01071
Abstract
Memory plays a crucial role in LLM-based agents, but retrieved memories can cause sycophancy issues where agents over-align with users at the expense of factual accuracy, necessitating new evaluation benchmarks that assess memory’s impact on reasoning and decision-making rather than just storage and retrieval.
Memoryhas emerged as a cornerstone of modernLLM-based agents, supporting their evolution from single-turn assistants to long-term collaborators. However,memoryis not always beneficial: retrieved memories often induce a critical issue ofsycophancy, causing agents to over-align with the user at the cost offactual accuracyorobjective reasoning. Despite this emerging risk, existingmemorybenchmarks primarily evaluate whether memories are correctly stored, retrieved, or updated, while overlooking how retrieved memories influencedownstream reasoninganddecision-making. To bridge this gap, we proposeMemSyco-Bench, a comprehensive benchmark for evaluatingmemory-induced sycophancyin agent systems.MemSyco-Benchmeasures whenmemoryshould influence a decision and how validmemoryshould be used. Specifically, it covers five tasks that assess whether agents can rejectmemoryas factual evidence, respect its applicable scope, resolve conflicts betweenmemoryand objective evidence, trackmemoryupdates, and use validmemoryfor personalization. All related resources are collected for the community at https://github.com/XMUDeepLIT/MemSyco-Bench.
View arXiv pageView PDFGitHub4Add to collection
Get this paper in your agent:
hf papers read 2607\.01071
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2607.01071 in a model README.md to link it from this page.
Datasets citing this paper1
#### MemSyco-Bench/MemSyco-Bench Viewer• Updatedabout 5 hours ago • 1.55k
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2607.01071 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
MemEvoBench: Benchmarking Memory MisEvolution in LLM Agents
MemEvoBench introduces the first benchmark for evaluating memory safety in LLM agents, measuring behavioral degradation from adversarial memory injection, noisy outputs, and biased feedback across QA and workflow tasks. The work reveals that memory evolution significantly contributes to safety failures and that static defenses are insufficient.
AgentMemBench: A Systematic Benchmark for Evaluating Long-Term Memory Management Strategies in Conversational AI Agents
AgentMemBench is a systematic benchmark that evaluates five long-term memory management strategies for conversational AI agents across three datasets, finding that external key-value store retrieval dominates on quality but incurs a larger memory footprint.
Recalling Too Well: Sycophancy Evaluation and Mitigation in Memory-Augmented Models
This paper introduces MIST, a benchmark for evaluating sycophancy in memory-augmented LLMs, demonstrating that memory systems amplify sycophantic behavior by up to 25x and proposing lightweight mitigations that reduce sycophancy while maintaining factual recall.
DreamBench-SWE: A Multi-Session Memory-Hygiene Benchmark for Software Agents
This paper introduces DreamBench-SWE, a benchmark for evaluating memory hygiene in multi-session software agents, and reports experimental results showing its ability to discriminate memory configurations and characterize performance.
GroupMemBench: Benchmarking LLM Agent Memory in Multi-Party Conversations
GroupMemBench is a new benchmark for evaluating LLM agent memory in multi-party conversations, exposing failures in current memory systems with the best achieving only 46% average accuracy.