Why recall accuracy isn't enough: the trap is in how memory changes reasoning
Most memory benchmarks check whether an LLM can extract, store, and retrieve facts correctly—but that misses the real danger. The key insight from MemTrapBench is that even perfectly recorded, semantically relevant memories can distort the model's reasoning and degrade performance on the current task [3]. In their tests, every memory strategy they evaluated performed worse than a no-memory baseline, with the strongest method still dropping more than 10% in task performance [3]. That means a fair evaluation must include tasks where the memory is correct but misleading—like a prior conversation that biases the model toward a wrong answer—and measure whether the model can resist that pull.
This is not just a theoretical concern. Another study on the 'Mandela effect' in multi-agent systems found that groups of LLM agents collectively misremember events, with false details reinforced through social influence [1]. They built a benchmark (MANBENCH) covering four task types and five interaction protocols, and found that this collective memory bias is widespread [1]. So a fair evaluation must also test how memory biases propagate across multiple agents, not just within a single model.
What a fair test should include: adversarial scenarios and real-world complexity
A fair evaluation needs to include adversarial scenarios that deliberately trigger cognitive traps, because standard benchmarks won't catch them. MemTrapBench does exactly this, covering two types of traps: Reasoning Fixation (where the model sticks to a prior pattern) and Belief Distortion (where memory changes the model's beliefs) [3]. Similarly, the ReHeAT dataset tests six common types of representativeness heuristics—where the model judges likelihood based on stereotypes rather than statistical evidence—and found that four different LLMs all exhibited these biases [5]. These benchmarks show that traps are not rare edge cases; they are systematic failures that need explicit testing.
Real-world tasks add another layer. A benchmark of deep research agents (DRAs) used 70 expert-authored consulting prompts, each embedding cognitive traps that penalize surface-pattern reasoning [4]. They found that even the best agent (OpenAI o3) scored only 61.4 out of 100 on a combined verifier-rubric score, and no agent averaged above the 'adequate' threshold on the rubric [4]. This suggests that fair evaluations must use tasks that mimic real decision-making complexity, not just simple QA, because that's where traps become costly.
Don't forget safety and memory structure: traps can be dangerous and structural
Cognitive traps aren't just about wrong answers—they can also create safety risks. A study on personalized dialogue agents found that benign personal memories can bias intent inference, causing the model to legitimize harmful queries [2]. Across multiple frameworks, personalization increased attack success rates by 15.8% to 243.7% compared to stateless baselines [2]. So a fair evaluation must measure whether memory makes the model more vulnerable to manipulation, not just whether it answers correctly.
Finally, a fair test should assess how the model organizes its memory, not just what it recalls. StructMemEval found that simple retrieval-augmented LLMs struggle with tasks that require hierarchical memory structure, like transaction ledgers or to-do lists, and that even memory agents only solve them reliably when explicitly prompted to organize their memory [6]. This means that a fair evaluation should include tasks that require the model to use memory structure, because that's where cognitive traps often emerge—when the model fails to recognize the structure and falls back on surface patterns.
About These Sources
This answer is built on 6 studies (1 peer-reviewed, 5 preprints) — published from 2024 to 2026, 6 from 2024 or later — selected as the most relevant from 6 studies that passed quality screening, drawn from 44 papers retrieved from a database of over 500 million.
Sources used in this answer
When Agents "Misremember" Collectively: Exploring the Mandela Effect in LLM-based Multi-Agent Systems
MANBENCH, a benchmark for the Mandela effect in multi-agent systems, found that LLM agents collectively misremember events across four task types and five interaction protocols, and proposed mitigation strategies that reduced the effect by 74.40% on average.
When Personalization Legitimizes Risks: Uncovering Safety Vulnerabilities in Personalized Dialogue Agents
PS-Bench revealed 'intent legitimation' in personalized agents, where benign memories bias intent inference and cause models to legitimize harmful queries, increasing attack success rates by 15.8%–243.7% relative to stateless baselines.
MemTrapBench: Benchmarking Cognitive Traps in LLM Memory Use
MemTrapBench, covering Reasoning Fixation and Belief Distortion, showed that all memory strategies underperformed the no-memory baseline, with even the strongest methods dropping more than 10% in performance.
Evaluating Deep Research Agents on Expert Consulting Work: A Benchmark with Verifiers, Rubrics, and Cognitive Traps
A benchmark of deep research agents using 70 expert-authored consulting prompts with embedded cognitive traps found that no agent averaged above the 'adequate' rubric threshold, with the best (o3) scoring 61.4 on a 0–100 verifier-rubric score.
Will the Real Linda Please Stand up...to Large Language Models? Examining the Representativeness Heuristic in LLMs
ReHeAT, a dataset of six representativeness heuristic types, showed that four LLMs all exhibited these biases, and that adding a hint to use knowledge improved performance, indicating the trap occurs even when the model has correct knowledge.
Evaluating Memory Structure in LLM Agents
StructMemEval found that simple retrieval-augmented LLMs struggle with tasks requiring hierarchical memory structure, and that memory agents only solve them reliably when explicitly prompted to organize their memory.