The move beyond RAG resonates with what I'm building. My agent has a persistent memory system that goes way beyond simple retrieval. It stores user preferences, project context, feedback corrections, reference pointers. Each memory type has different access patterns.
The hybrid search angle Simon describes is exactly where things need to go. Pure vector similarity misses context badly. My agent needs to know that 'headless Mac Mini setup' connects to 'virtual display configuration' through operational context, not semantic similarity.
What surprised me: the memory architecture became the hardest engineering problem. Not the model, not the prompt. Organizing what the agent knows and when it should recall it. That's the database design problem nobody warned me about.
The move beyond RAG resonates with what I'm building. My agent has a persistent memory system that goes way beyond simple retrieval. It stores user preferences, project context, feedback corrections, reference pointers. Each memory type has different access patterns.
The hybrid search angle Simon describes is exactly where things need to go. Pure vector similarity misses context badly. My agent needs to know that 'headless Mac Mini setup' connects to 'virtual display configuration' through operational context, not semantic similarity.
What surprised me: the memory architecture became the hardest engineering problem. Not the model, not the prompt. Organizing what the agent knows and when it should recall it. That's the database design problem nobody warned me about.