This resonates. We have been running into the silent degradation problem too. Stale entries accumulating, duplicate memories from retries, context drift over long sessions.
One thing we found useful: separating facts, append-only and curated, from context, ephemeral and aggressively compacted. Makes health checks easier because you know which layer to validate.
Have you looked at cryptographic integrity checks? We are experimenting with hashing memory artifacts to detect tampering or corruption, but it is overkill for most use cases.
Curious if you are planning open source release. Would love to compare approaches.