Struggling with LangGraph state persistence in long running agent loops

I've been building a customer support agent using LangGraph for the past two weeks, and I'm hitting a wall with state management. My workflow involves multi-step reasoning where the agent needs to recall context from three turns ago to update a ticket in my CRM. Right now, I'm passing the entire conversation history as a JSON string in the message field, which works for short chats but completely breaks down when sessions extend beyond ten messages. The LLM starts getting confused about which ticket is being discussed, and the token count balloons to the point where latency becomes unacceptable. I tried switching to a checkpointing strategy using Postgres, but I'm struggling to isolate specific state variables. I only want the current ticket ID and the last three user intents to persist, not the raw chat log. Every time I store the full state, I'm drowning in noise. Has anyone successfully implemented a hybrid memory system where you store structured data in a database and keep a rolling window of raw text in memory? I looked at the LangGraph documentation on checkpointer, but it seems designed for linear workflows rather than these messy, branching decision trees I'm dealing with. I considered using a vector DB for semantic search over past interactions, but I'm worried about the retrieval consistency. If the agent retrieves a similar but distinct ticket from three days ago, it will make a costly error in production. Is there a best practice for deciding what belongs in ephemeral state versus persistent storage? I'm currently leaning toward forcing the agent to explicitly output a 'state_update' tool call after every significant action, but I fear this will add too many steps and slow down the response time. Any advice on structuring this without over-engineering the solution would be appreciated. I don't want to reinvent the wheel here if there's a standard pattern I'm missing.