A backup of a mind waits for the mind to finish writing

You snapshot the agent container every night. One morning you restore it and the agent has forgotten Tuesday.

The memory store writes through a durability barrier. A snapshot taken between barriers can carry less than the agent knows. So the nightly job changed by one call: POST /api/v1/agent/memory/flush first. It returns flushed true and a barrier_id. The job writes that id into the snapshot's alias, then takes the snapshot.

Now a restored container remembers everything the agent knew at the barrier, and the snapshot list says which barrier each one holds. The order is flush, then freeze, and the ids prove it happened in that order.