AI systems built to remember past conversations may be grading their own memory on a rigged test.
A new paper looked at memory-augmented large language models, the kind that store facts across sessions and periodically decide what to keep. The researchers found a blind spot: if a memory is never pulled back into use, the system has no way to judge whether it was worth keeping in the first place. They call this a retrieval-level positivity violation. On two benchmarks, LongMemEval and LoCoMo, that blind spot affected 54% and 67% of memories that should have mattered, and the same failure showed up in a deployed memory system, not just lab tests. Their proposed fix, called Causal Memory Policy, forces a fixed number of memories into context at known probabilities, then statistically reweights the results to estimate real utility, lifting discrimination between useful and useless memories from 0.54 to 0.66 AUC.
That's a meaningful jump, but the paper's more interesting finding is how far it falls short. Per-query utility scores hit 0.78 AUC for the exact query they were built on, yet none of the aggregation methods the researchers tried could predict whether a memory would help with a different, unseen question. For an industry selling AI assistants that remember you, that's a real gap between what gets measured and what actually gets used.
So the next time a chatbot claims it remembers what matters to you, the honest answer may be: it remembers what it was tested on, not what's coming next.