Can relevant memories actually harm LLM reasoning?

The paper's claim is that memory can fail at the point of use, after extraction, storage and retrieval have all worked. It calls these failures "memory-induced cognitive traps": "even faithfully recorded and semantically relevant memories can distort model reasoning or beliefs and degrade current task performance." The introduction frames the gap directly. Existing benchmarks "primarily assess memory extraction, storage, updating, and retrieval," while "largely overlooking how retrieved memories reshape model reasoning." MemTrapBench covers two forms of trap, Reasoning Fixation and Belief Distortion. Across two model families and five representative memory frameworks, "all evaluated memory strategies underperform the no-memory setting," and even the strongest suffer drops of more than 10%.

The reasoning is a change of evaluation target. If the only questions asked are whether the right item was extracted, stored and retrieved, a memory system can score well while making the model worse at the task in front of it. The paper's own phrase is "memory is not always what we need, as it may impair rather than enhance model capabilities." The proposed mitigation, AdaptiveMem, sits at the same point of use. It is an inference-time method that "instructs LLMs to avoid memory traps," and is reported to mitigate the traps on MemTrapBench while preserving or improving performance on standard memory benchmarks across diverse memory frameworks.

This lands next to two existing notes and extends both. Do memory systems actually help language models learn continuously? also finds memory systems falling below a stateless baseline, but locates the cause in stale beliefs and spurious generalizations built up over sequential experience. The MemTrapBench claim is narrower and harder to dismiss, because it concerns memories that are faithfully recorded and relevant, so staleness or a corrupted store cannot be the explanation. Does retrieved memory quality depend on its functional role? shows irrelevant memory hurting. This paper says passing the relevance test is not a sufficient guard either, so relevance filtering alone would not remove the failure it describes.

The excerpt leaves most of the evidence out. It does not name the models or the five frameworks, describe how Reasoning Fixation and Belief Distortion are constructed or scored, say which metric the ">10%" refers to, or give the size of AdaptiveMem's gains. The discussion passage only restates that the benchmark "remains challenging." Nor does it say whether an instruction-level fix holds beyond the benchmarks it was tested on. What the excerpt does support is a modest methodological point: a memory evaluation without a no-memory baseline cannot show that memory helps, and retrieval accuracy is not evidence that the retrieved memory is being used well.