AuthorsMengru Wang, Haozhe Luo, Zhenqian Xu et al.
TL;DR
MemTrapBench exposes memory-induced cognitive traps and shows AdaptiveMem’s prompt-based guardrail recovers up to 14.9 percentage points on long-horizon memory benchmarks.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
MemTrapBench shows that even the strongest memory methods suffer drops of more than 10 percentage points compared to the no-memory setting.
On Gemini-3-Flash-Preview and Qwen3-30B-A3B-Instruct-2507, memory frameworks like LightMem and EverMemOS trigger Reasoning Fixation and Belief Distortion, degrading current-task performance.
HOW IT WORKS
MemTrapBench builds on a Taxonomy, Instance Construction, and Two-Gate Quality Flow to generate 1,050 trap-inducing multi-turn dialogues across four scenarios.
You can think of MemTrapBench like a stress lab where past interactions act as a misleading "training set" that biases a live reasoning engine, while AdaptiveMem behaves like a safety inspector.
This design lets MemTrapBench expose how memory reshapes reasoning strategies in ways a plain context window benchmark cannot, and how AdaptiveMem selectively neutralizes harmful memories without discarding useful ones.
DIAGRAM
This diagram shows how MemTrapBench plants, buries, and then springs a cognitive trap within a multi-turn interaction.
DIAGRAM
This diagram shows how MemTrapBench evaluates memory strategies, no-trap controls, and memory-length ablations.
PROCESS
Taxonomy
MemTrapBench first assigns each seed to Reasoning Fixation or Belief Distortion using the Taxonomy, covering Cognitive Bias, Task Boundary, Trauma, and Safety.
Instance Construction
MemTrapBench performs Instance Construction by expanding seeds into multi-turn dialogues that Plant the Trap, Bury It in Noise, and Spring the Trap.
Two-Gate Quality Flow
MemTrapBench runs the Two-Gate Quality Flow, combining AI-based filters and Human-Expert Review to enforce topic coherence, context consistency, and standalone solvability.
AdaptiveMem
MemTrapBench then evaluates memory strategies with and without AdaptiveMem, which prompts LLMs to detect potential traps before using retrieved history.
KEY CONTRIBUTIONS
Memory-induced Cognitive Traps
MemTrapBench formally defines Memory Traps where s(ŷM) < s(ŷ∅) and shows that all evaluated memory strategies underperform the no-memory baseline by more than 10 percentage points.
MemTrapBench Benchmark
MemTrapBench releases 1,050 adversarial instances across Cognitive Bias, Task Boundary, Trauma, and Safety, generated via the Taxonomy, Instance Construction, and Two-Gate Quality Flow.
AdaptiveMem
MemTrapBench introduces AdaptiveMem, a prompt-based skill that improves LightMem on Gemini-3-Flash-Preview by 14.9 percentage points on MemTrapBench without degrading LongMemEval performance.
RESULTS
Avg. score Gemini wo/Mem
85.16%
+14.0 over FullText on MemTrapBench
Avg. score Gemini EverMemOS
71.17%
+0.06 over LightMem on MemTrapBench
Avg. score Qwen wo/Mem
81.83%
+10.84 over LightMem on MemTrapBench
Trauma no-trap correctness
91.07%
+24.67 over MemTrapBench Trauma setting
These metrics come from MemTrapBench’s main results tables, which evaluate Gemini-3-Flash-Preview and Qwen3-30B-A3B-Instruct-2507 across Cognitive Bias, Task Boundary, Trauma, and Safety. The numbers show that MemTrapBench reliably induces performance drops when memory is used, and that removing trap cues or adding AdaptiveMem can restore correctness and efficiency.
BENCHMARK
These metrics come from MemTrapBench’s main results tables, which evaluate Gemini-3-Flash-Preview and Qwen3-30B-A3B-Instruct-2507 across Cognitive Bias, Task Boundary, Trauma, and Safety. The numbers show that MemTrapBench reliably induces performance drops when memory is used, and that removing trap cues or adding AdaptiveMem can restore correctness and efficiency.
BENCHMARK
Average score (%) across all MemTrapBench scenarios for Gemini-3-Flash-Preview with different memory strategies.
KEY INSIGHT
MemTrapBench shows that every evaluated memory strategy on Gemini-3-Flash-Preview scores below the 85.16% no-memory baseline, with drops exceeding 10 percentage points.
This is surprising because memory frameworks like LightMem and EverMemOS are designed to help long-term reasoning, yet MemTrapBench reveals that faithfully retrieved history can systematically hurt performance.
WHY IT MATTERS
MemTrapBench gives practitioners a concrete way to stress-test how external memory reshapes reasoning, rather than just checking extraction and retrieval accuracy.
With MemTrapBench and AdaptiveMem, builders can now design agents that benefit from long-term memory while explicitly guarding against fixation, feedback-induced avoidance, and safety belief distortions.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Shannan Yan, Jingchen Ni et al.
· 2026
AdaMem organizes dialogue history into Working Memory, Episodic Memory, Persona Memory, and Graph Memory coordinated by a Memory Agent, Research Agent, and Working Agent. On LoCoMo with GPT-4.1-mini, AdaMem achieves 44.65 F1 overall, beating the best baseline LangMem at 41.76 F1 by +2.89.
Answers use this explainer on Memory Papers.
Checking…