LangGraph state persistence failing during long-running multi-agent workflows

I've been wrestling with a persistent state management issue in my LangGraph implementation for the past two days. I'm building a research assistant that splits work between a planner agent and a verifier agent. The planner generates a detailed outline, then hands control to the verifier to check for hallucinations before finalizing the report. The problem occurs when the workflow exceeds roughly five minutes. I'm using Redis for state persistence, which should be fine for short interactions, but for these deep-dive research tasks, the graph seems to lose the intermediate state of the 'verification' node. When the verifier finishes and tries to route back to the planner for corrections, the planner acts as if it has no memory of the original outline, leading to circular loops where it regenerates the same flawed plan. I've checked the Redis connection logs, and there are no timeouts or disconnections. The TTL is set to 24 hours, so it's not expiring. I initially thought it was a serialization issue with the Pydantic models I'm passing through the state, so I tried switching to JSON strings, but that made it worse because the agents couldn't parse the context correctly. Has anyone else hit this wall with long-running LangGraph chains? I'm currently considering breaking the workflow into smaller, synchronous chunks and manually managing the handoff via a simple API endpoint, but that feels like a step backward in terms of autonomy. I'd prefer to keep the graph logic intact. Is there a recommended pattern for handling state checkpoints that survive extended execution times? Or am I better off switching to a different framework like CrewAI for this specific type of collaborative task? Any insights on debugging the state graph specifically would be appreciated.