Found first: a primary source the press has not covered yet.
A new benchmark tests whether LLM memory modules degrade model performance even when the memories retrieved are accurate and relevant. The paper MemTrapBench: Benchmarking Cognitive Traps in LLM Memory Use tests five representative memory frameworks against two model families and finds every framework underperforms a no-memory baseline, with the best methods falling more than 10 percentage points below baseline scores.
What the source says
The benchmark identifies two failure pathways: Reasoning Fixation, in which memories lock models into flawed inference chains across three sub-types (Cognitive Bias, Trauma, and Task Boundary), and Belief Distortion, in which memories corrupt model beliefs about current reality. Testing Gemini-3-Flash-Preview (85.16% baseline) and Qwen3-30B-A3B-Instruct-2507 (81.83% baseline) against five frameworks, FullText, LightMem, MemOS, SimpleMem, and EverMemOS, all methods fell to around 70-71% at best. On Cognitive Bias scenarios specifically, Gemini's performance with memory ranged from 46.66% to 65.48%. The authors release AdaptiveMem, a prompt-based inference-time method that raised LightMem's score by 14.9 percentage points on MemTrapBench while preserving performance on the standard LongMemEval benchmark.
Why it matters
Memory modules are a standard component in agent products, and evaluation has focused almost entirely on retrieval accuracy, treating correct recall as a proxy for downstream correctness. This benchmark finds the failure is structural: accurate, relevant memories still introduce reasoning shortcuts and context contamination that current architectures do not handle. Every framework tested fails, which means the problem is not in any single implementation. Teams building agent systems that accumulate memory across sessions should test specifically for these failure modes, separate from retrieval correctness.