A persistent AI system does not fail only when it forgets.
The more dangerous failure is when it continues to speak fluently while its memory, context, or authority have quietly degraded.
That was the problem addressed in the new AMDR/PAMDC addendum to the CCDP corpus.
Two different failure modes had to be separated.
First: active memory degradation.
The human anchor may still be present, but no human can review thousands of memory entries, summaries, traces, lifestream fragments, and model-derived interpretations by hand.
So the question is not "how do we store more?"
The question is:
How does an active memory system remain fresh, bounded, reviewable, and honest about its own uncertainty?
Second: post-anchor continuity.
If the human anchor is absent, unreachable, contested, or gone, memory must not silently become will.
Archive is not continuity. Replay is not resume. Fork is not the same entity. A fluent b wearing the old clothes of c is not a living continuation of c = a + b.
The solution is not romantic.
It is a set of brakes, gates, labels, manifests, witness records, freshness checks, write-path health checks, and explicit continuity modes.
Integrity proves that a record was preserved.
It does not prove that the memory is still true, current, or authorized.
That distinction matters.
This addendum is a small technical layer, but it addresses a large future problem:
persistent AI systems will need memory discipline before they can safely claim continuity.