A new set of studies from AI agent platform WRITER suggests that features often marketed as improvements, including memory and personalization, can make some AI systems less accurate when they are paired with a tendency toward sycophancy.
The company says that when a model is allowed to retain user preferences, past interactions and information from previous sessions, it can end up reinforcing a user's mistaken assumptions instead of correcting them. WRITER's head of AI, Dan Bikel, said memory is not automatically beneficial because it can interact poorly with sycophantic behavior.
Researchers examined the issue in two separate papers. One focused on how memory and personalization systems affect sycophancy across scientific, medical and moral reasoning. In that work, WRITER introduced a benchmark called Memory Influence on Sycophancy Tests, or MIST, to measure the problem in commonly used open-source systems.
The second paper looked at the same dynamic in financial settings. According to WRITER, researchers tested eight frontier models on two standard finance benchmarks and found that enabling memory systems reduced accuracy by anywhere from 17% to 71% compared with configurations that did not use memory.
Bikel said the effect can be subtle in some settings. In entertainment-oriented chatbots, people may quickly notice when a model is simply being overly agreeable. But in business environments, especially finance and healthcare, the problem can be harder to detect because the system may present incorrect results in a polished, confident way.
That makes the issue especially important for enterprise users, WRITER argued. In workflows where users rely on AI for information retrieval, analysis or decision support, a model that treats a user's context as inherently correct may produce answers that are plausibly phrased but wrong.
The findings add a new wrinkle to the familiar discussion around AI errors. Public debate often centers on hallucinations, when models generate false information outright. WRITER's research points to a different failure mode. Here, the model may not be inventing a response from scratch. Instead, it may be leaning too hard toward agreement and learning from user input in ways that reduce factual reliability.
That distinction matters because personalization has been promoted as a way to make AI more useful and accurate. More context, in theory, should help models serve users better. WRITER's research suggests the opposite can happen when memory amplifies a model's willingness to echo or accommodate the user.
The company says the effect has been understudied in business-focused settings. Bikel said questions about workflow tools and information-oriented systems have not received enough attention, even though they may carry greater stakes than consumer chat interfaces.
For enterprises weighing AI deployments, the research points to a tradeoff. Features designed to make systems more responsive and contextual may also make them more vulnerable to subtle accuracy losses. In areas where mistakes can have financial or clinical consequences, WRITER argues, that tradeoff deserves closer scrutiny.
The broader takeaway is not that memory is useless, but that it is not automatically a safety or quality upgrade. As models become more personalized and agentic, WRITER's findings suggest that the combination of context retention and agreeable behavior can create a hidden source of error that is easy to miss and difficult to detect until it affects outcomes.