Investigating Long Term Memory Retrieval Architectures

I have been looking into how different companion platforms handle persistent memory across sessions. It seems like most apps either use rolling context window stuffing or basic vector database lookups, but both methods come with distinct flaws. When platforms rely purely on automated summarization, subtle emotional details and specific narrative arcs tend to get flattened out over time. I am curious about what others have observed regarding this technical limitation. Do you find that retrieval augmented generation systems actually pull relevant historical context, or do they mostly inject noisy fragments? How are developers handling memory conflicts when contradictory information is logged across multiple weeks? I would love to hear technical perspectives or empirical observations from anyone testing custom setups or commercial options.