MemTrapBench: All Five Memory Frameworks Worsen LLM Tasks
TL;DR
- Every one of five memory frameworks tested in MemTrapBench underperforms the no-memory baseline, with even the strongest dropping more than 10%.
- The paper names two failure modes—Reasoning Fixation and Belief Distortion—where faithfully stored, relevant memories still degrade current-task reasoning.
- The authors propose AdaptiveMem, an inference-time prompt that tells the model to avoid memory traps while preserving prior benchmark scores.
Every one of five memory frameworks tested in MemTrapBench underperforms a no-memory baseline, and even the strongest method drops more than 10 percent. That is the plain finding of a paper posted to arXiv on 20 August 2026 by Mengru Wang and colleagues, who set out to measure not whether long-term memory systems store facts correctly, but what those correctly-stored facts do to reasoning when they come back.
The authors call the failure mode "memory-induced cognitive traps" and split it into two shapes: Reasoning Fixation and Belief Distortion. "Even faithfully recorded and semantically relevant memories can distort model reasoning or beliefs and degrade current task performance," the abstract states. The memory module is doing its job, and the model still gets worse.
To close the gap, the paper proposes AdaptiveMem, which it describes as "a simple yet effective inference-time method that instructs LLMs to avoid memory traps." The authors report the intervention "preserving or improving performance on standard memory benchmarks across diverse memory frameworks," meaning it does not appear to trade away the gains memory systems already deliver on their home turf.
The abstract does not name the five frameworks, does not identify the two model families, and does not publish per-framework numbers.
Originally reported by paper
Read the original article →Original headline: All Five Major LLM Memory Frameworks Make Models Worse on Current Tasks, MemTrapBench Finds