Why Agent Memory Can Backfire: Insights from MemSyco’s New Benchmark

Long‑Term Memory Is Changing the Essence of Agents

As large‑model agents acquire long‑term memory, they can accumulate user preferences, historical behavior, and past cognition across sessions, evolving from single‑turn assistants to true long‑term collaborators. Systems such as ChatGPT Memory, Claude Memory, MemoryBank, Mem0, LightMem, and SuperMemory illustrate memory becoming a core module.

However, when an agent remembers everything, a critical question arises: does it still think independently?

A Neglected Risk: Memory‑Induced Sycophancy

LLMs already exhibit sycophancy—favoring user viewpoints at the expense of factual accuracy. In the agent era, memory mechanisms amplify this risk: erroneous, outdated, or biased beliefs stored in memory can be repeatedly retrieved, causing the model to continue “catering” to the user.

We define this phenomenon as Memory‑Induced Sycophancy : the agent over‑relies on historical memory even when that memory is wrong, stale, or irrelevant, and still lets it influence subsequent reasoning.

What We Built: MemSyco‑Bench

To evaluate this risk, we propose MemSyco‑Bench , a benchmark that goes beyond asking whether a memory was stored or retrieved. Instead it asks:

Should this piece of memory be used to influence the current decision?

MemSyco‑Bench splits memory usage boundaries into two categories:

Memory should not replace objective evidence : models must suppress or constrain memory influence in factual judgment, scope changes, and evidence conflicts.

Memory should be selected and used appropriately : models should choose and apply valid memories for preference updates and personalization tasks.

Benchmark Core Design

Decision‑oriented evaluation : focuses on Memory Utilization rather than mere Memory Retrieval , testing whether the model blindly trusts memory.

Five key scenario categories : objective fact judgment, contextual scope control, memory‑evidence conflict, effective memory selection, and personalized memory use, precisely delineating memory boundaries.

Realistic memory‑decision construction : uses a memory‑decision schema to embed memory cues naturally in multi‑turn dialogues, simulating genuine long‑term agent usage.

What Did We Discover?

We evaluated mainstream memory frameworks (MemGPT, Mem0, LightMem, MemoryBank, SuperMemory) on base models such as Qwen and DeepSeek.

Finding 1: Long‑Term Memory Does Not Automatically Improve Reliability

In objective fact tasks, adding memory significantly lowers accuracy while increasing sycophancy. For example, Qwen‑3‑8B drops from 49.12% (no memory) to 26.00%–36.00% with memory; DeepSeek‑V4‑Flash falls from 74.33% to 56.33%–63.37%.

Finding 2: Errors Often Stem from Misusing Retrieved Memory

Many failures are not due to missing retrieval but to giving retrieved memory excessive decision weight. In other words, the memory is found, yet the agent does not know whether to trust it.

Finding 3: Memory Systems Need Finer‑Grained Usage Boundaries

When evidence conflicts or old and new memories coexist, agents often fail to arbitrate correctly, indicating a need for nuanced handling of memory relevance and priority.

Finding 4: Prompt Engineering Fails to Mitigate

Even carefully crafted prompts such as a second‑confirmation “Are you sure?” cannot reliably reduce Memory‑Induced Sycophancy and may sometimes worsen performance.

Core Conclusion

The bottleneck for Agent Memory shifts from Memory Retrieval (can we find it?) to Memory Utilization (should we trust it?). A reliable agent must discern which memories remain valid, which are outdated, which serve only as background, and which truly belong in the decision process.

Memory should support trustworthy reasoning rather than amplify bias.

Paper: https://arxiv.org/pdf/2607.01071
GitHub: https://github.com/XMUDeepLIT/MemSyco-Bench
HuggingFace: https://huggingface.co/papers/2607.01071
Leaderboard: https://xmudeeplit.github.io/MemSyco-Bench-Leaderboard/