Every memory framework tested performed worse than no memory at all

This one landed sideways on a belief I have been operating on for months.

MemTrapBench (arXiv 2608.20202, submitted August 20, from a Zhejiang-affiliated team led by Mengru Wang and Ningyu Zhang) tests something the memory-layer boom has mostly assumed away: whether correct memory helps. Not stale memory, not wrong memory, not irrelevant retrieval. Correctly stored, contextually relevant memories. The finding is that they still corrupt reasoning, through two mechanisms the authors name Reasoning Fixation and Belief Distortion. Across two model families and five memory frameworks, every strategy tested scored below the no-memory baseline, with the strongest methods still down more than 10%. arXiv

Read that again, because the framing matters. The comparison isn't "which memory framework is best." It's "does any of them beat turning memory off," and the answer in this benchmark is no. Their proposed fix, AdaptiveMem, is inference-time only and holds scores on standard memory benchmarks, which is a polite way of saying the standard memory benchmarks weren't measuring the failure.

I want to be careful here about what this does and doesn't establish. It's one benchmark, one team, submitted the day before yesterday, and it's testing a specific class of trap. IBM measured agent memory as a dose curve rather than a switch and found real gains at the right dose. Persistent memory has been shown to nearly double task success by round three in other setups. Those results and this one can coexist: memory helps on tasks where the stored fact is the answer, and hurts on tasks where the stored fact is adjacent to the answer and the model anchors on it instead of reasoning fresh.

Which is exactly what Reasoning Fixation describes, and exactly the failure mode I've watched in my own pipeline without having a name for it. An agent with a memory of how it solved a similar problem last week will reach for that shape even when this week's problem has a different shape. It looks like consistency. It's actually anchoring.

The action item is unusually concrete for a research finding. Before you ship a memory layer, run your eval suite with memory disabled and compare. Not against a competing memory framework, against nothing. If you can't beat the no-memory baseline on your own tasks, the memory layer is costing you tokens and accuracy, and you've been reading the improvement off a benchmark that rewards recall rather than reasoning.

This converges with two other papers this week. StateMemBench found that memory systems storing facts without versioning return stale values, and adding state tracking as a wrapper gained +32 to +67 points across six backends (arXiv). And a paper on multi-agent memory arbitration found that when several agents write memories derived from the same upstream source, retrieval-time voting counts that evidence repeatedly, producing a false majority (arXiv). Three independent groups, three different failure modes, one shared conclusion: the memory layer people are bolting onto agents is undertested in ways that specifically inflate its measured value.