What's this about?
A benchmark exposing a hidden flaw in AI long-term memory systems: they assume a stored memory will look similar to the question that later needs it. When the connection is implicit rather than word-for-word, agents fail to retrieve the right memory.
What exactly is "agent memory"?
To remember things across conversations, an agent saves facts you tell it into an external store, then pulls them back when a related question comes up. Retrieval usually works by matching similar wording. The blind spot: if you once said "I'm allergic to shellfish" and later ask "can I eat this shrimp dish?", a naive memory system may not connect them because the words don't overlap — the association is implicit.
Why is this worth noting?
Memory is central to useful personal agents, and this names a failure mode that lexical-similarity retrieval quietly produces. Benchmarks that isolate a specific weakness tend to drive the next round of fixes.