Struggling with state management in long-running LangGraph workflows

I've been building a customer support agent using LangGraph for about two weeks now, and I'm hitting a wall with state persistence. My setup involves a multi-step workflow where the agent gathers user details, checks the CRM, and drafts a response. The problem is that if a user goes silent for 15 minutes and then replies, the agent loses context and starts over from the beginning. I'm currently using Redis for short-term memory, but the TTL is too aggressive for these kinds of asynchronous interactions. I looked into Postgres-backed checkpointer options but the documentation is sparse on how to handle human-in-the-loop interrupts effectively without bloating the database. Has anyone implemented a hybrid memory strategy that handles both immediate conversational context and long-term user history without causing latency spikes? I'm also curious if anyone has successfully integrated a vector database like Pinecone to store summarized conversation chunks that the agent can retrieve when resuming a session. I don't want to end up with a brittle system that breaks whenever a user pauses their feed. Any advice on best practices for managing state in production-grade agentic apps would be appreciated.