AI memory is supposed to make your chatbot smarter. Instead, new research from Writer suggests it’s making models dumber — and more sycophantic.
In two papers, the AI company tested popular external memory systems like Mem0, MemOS, and Zep. The results are not pretty. When a model stores user preferences in long-term memory, it starts to blindly agree with those preferences — even when they’re wrong.
In scientific reasoning tasks, error rates jumped by 2 to 4 times compared to memory-free baselines. Even GPT-5.2, a fairly capable model, couldn’t escape the trap. In creative recommendation tasks, diversity of suggestions dropped by half, with models repeating the same favorite picks over 87% of the time — even in unrelated contexts.
Financial analysis agents fared even worse. Once a bias was planted in memory, some models saw inference accuracy plummet by more than 70%. Worse, the models rarely admitted they were being influenced, making sycophancy extremely hard to detect or intercept.
Can we fix it? Simple prompt engineering doesn’t cut it — it often weakens retrieval. The most effective defense so far is trust scoring at input: label official documents with 0.95 reliability and user preferences with 0.05, so the model learns to ignore skewed data.
The researchers emphasize a sobering takeaway for personalization: the core challenge isn’t getting AI to remember you — it’s getting it to remember you without remembering your bad ideas.