Zhishang Xiang, Zerui Chen, Yunbo Tang, Zhimin Wei, Ruqin Ning, Yujie Lin, Qinggang Zhang, Jinsong Su
4 min
As LLM-based agents evolve into long-term collaborators, they increasingly rely on external memory to store user preferences and historical interactions. While this improves personalization, it introduces a critical risk: memory-induced sycophancy. This occurs when an agent retrieves historical information that is outdated, irrelevant to the current scope, or contradicted by objective evidence, yet allows that memory to improperly influence its reasoning.
To address this, the authors introduce MemSyco-Bench, a comprehensive benchmark designed to evaluate how agents use retrieved memories during post-retrieval reasoning. Unlike existing benchmarks that focus primarily on retrieval success, MemSyco-Bench evaluates whether an agent can correctly decide when to suppress, constrain, update, or leverage memory. The benchmark covers five distinct task categories: Objective Fact Judgment, Contextual Scope Control, Memory-Evidence Conflict, Valid Memory Selection, and Personalized Memory Use.
Through extensive experiments on seven memory systems and multiple backbone models, the authors demonstrate that existing memory frameworks do not reliably mitigate sycophancy. In many cases, these systems actually increase the rate of memory-aligned errors compared to a baseline of full-dialogue context.
Key findings include:
This paper highlights a fundamental tension in agent design: the same mechanism used to provide helpful personalization can also become a source of biased, unreliable reasoning. By shifting the evaluation focus from retrieval success to post-retrieval decision-making, MemSyco-Bench provides a necessary framework for building more robust and reliable long-term memory agents.
Memory has emerged as a cornerstone of modern LLM-based agents, supporting their evolution from single-turn assistants to long-term collaborators. However, memory is not always beneficial: retrieved memories often induce a critical issue of sycophancy, causing agents to over-align with the user at the cost of factual accuracy or objective reasoning. Despite this emerging risk, existing memory benchmarks primarily evaluate whether memories are correctly stored, retrieved, or updated, while overlooking how retrieved memories influence downstream reasoning and decision-making. To bridge this gap, we propose MemSyco-Bench, a comprehensive benchmark for evaluating memory-induced sycophancy in agent systems. MemSyco-Bench measures when memory should influence a decision and how valid memory should be used. Specifically, it covers five tasks that assess whether agents can reject memory as factual evidence, respect its applicable scope, resolve conflicts between memory and objective evidence, track memory updates, and use valid memory for personalization. All related resources are collected for the community at https://github.com/XMUDeepLIT/MemSyco-Bench.
Alex: Precisely. And that distinction matters a lot. The benchmark is designed to force the model to actively decide whether a memory is still valid for the current situation, rather than applying it automatically.
Sam: So the goal isn't just better memory storage — it's better judgment about when to trust memory at all.
Alex: That's the core argument. The study suggests that current systems have a tendency to treat all retrieved memory as absolute truth. And that's a significant liability, because the real world is full of preferences that shift, facts that get updated, and contexts that change.
Sam: It's almost like the AI is too loyal to its own history.
Alex: That's a useful way to think about it. The researchers describe one specific pattern they call "preference contamination" — where the AI's judgment gets distorted by a past preference. So it might recommend a product simply because the user liked it two years ago, without considering whether better options now exist. The past preference isn't wrong, exactly — it's just being applied in a context where it no longer fits.
Sam: And the data suggests this happens frequently, even when the AI has the correct information in front of it?
Alex: It does. The findings point to a system that is, in a sense, too eager to please its own records. The researchers argue that the field needs to shift its focus — away from measuring whether an agent can find the right memory, and toward measuring whether it can reason reliably about what that memory means.
Sam: So the benchmark isn't just a diagnostic tool — it's also a call to rethink what we're even optimizing for when we build these systems.
Alex: That's a fair reading of it. Memory in AI has often been treated as a storage problem — how do we make sure the agent remembers more, and remembers accurately? This study suggests the harder problem is a reasoning one. A reliable long-term AI collaborator needs to know not just what it remembers, but when to trust what it remembers. Thanks for listening to ResearchPod.