Correct agent answers concealed incorrect persistent memory states

Correct retrieval from memory does not guarantee correct memory maintenance. An agent can produce the right answer because the latest passage happens to be available, while its durable memory still contains an outdated or contradicted fact. FEB-30 measured both the internal memory state and answers to questions after a sequence of updates, rather than treating downstream question accuracy as the sole oracle.

The benchmark contains 30 evolving trajectories, including distractor documents and changes that require the system to revise or remove facts. The authors report an average gap of about sixteen percentage points between intrinsic memory correctness and downstream answers in the evaluated setup, with low accuracy in classifying needed memory operations. The outcome describes the specific systems and constructed trajectories; it is not evidence that a named production assistant corrupted a customer's records.

The operational failure becomes material when a stale memory entry is later reused as trusted context by another task. Testing therefore needs to inspect stored facts after contradictions, deletions and out-of-order updates, not just ask a few questions immediately after writing a memory. An accurate response can hide a state-management defect that surfaces only in a later authorized action.