I know, you can find the answer in Notes on low-level discovery 

But I'm not agree with this feature :

{quote}

If an application is not discovered anymore all discovered items are automatically removed from it, even if the application itself is not yet removed because of the 'lost resources period' setting.

{quote}

@Ingus Vilnis

What do you think about ?

In my environment I have Cisco APs monitored base on Cisco WLC discovery. I've created a map with all APs and elements are the same Cisco WLC but with Application filter, if one is missing you cannot filter by application, and you cannot see alarms (using nodata() function since v3.4).

I suppose you can move 

/* add for removal the old links that aren't discovered and can be removed */
if (0 == (item_application->flags & ZBX_FLAG_LLD_ITEM_APPLICATION_DISCOVERED) &&
    FAIL == lld_item_application_validate(item_application, items, applications))
{
  zbx_vector_uint64_append(&del_itemappids, item_application->itemappid);
}

from lld_items_applications_save to lld_remove_lost_items

**It is by design.

I understand your issue as well. You have an environment with multiple APs. These access points are available through SNMP tree via Cisco WLC. Once the access point is not available it disappears from the SNMP tree. Quite the same approach is used to monitor processes on Windows computer via SNMP. It is really a problem. The indexes are changing as well. The only thing I can imagine is to swap the discovery with an external script and use snmpget + bash/python to use some cache mechanism and report the LLD JSON to suit our need.

Please register a feature request and introduce the logic to handle this case.

I've implemented with an external script, but I'm scary out performance. External scripts use a lot of CPU!