LLM Vector Memory Forgets Like Human Brains, Study Finds

Researchers spent several months measuring how AI agent memory degrades in vector stores and found that large language model memory systems follow the same mathematical forgetting patterns as human memory, mirroring results from classic clinical psychology experiments. The root cause lies in the low effective dimensionality of embeddings: despite advertising 384 to 1,024 dimensions, most models concentrate useful variance into roughly 16 effective dimensions, causing new memories to crowd and interfere with existing ones. This competition between stored items — not the passage of time — drives the forgetting curve, and removing competing memories reduced the forgetting exponent by around 50 times. The study also found that vector stores reproduce human-like false-memory rates, returning semantically plausible but never-stored facts at a rate of approximately 0.583, closely matching the 0.55 rate observed in classic human false-memory experiments. The practical implication is that scaling a vector index worsens recall rather than improving it, since more documents mean more competitors within a geometrically constrained embedding space.