Compared pure vector vs. hybrid retrieval and added a memory layer to my AI agent

First, some background. I'm currently working on a client project where the core task is having AI agents handle a batch of cross-departmental after-sales tickets. After running it for two weeks, I encountered a very typical problem: questions the agent answers correctly today get answered incorrectly tomorrow if phrased differently. It's not that the model is bad; it's that it lacks memory. Key information from the previous conversation turn—like customer IDs, device models, or previously promised resolution timelines—gets lost in the next turn.