Restarting AI Agents Requires Storing Plans, Pending Actions, and Memory Separately

A developer comment on dev.to by Adam Lewis asked where an agent's intermediate state lives when a terminal session closes, noting that saving the transcript is easy but preserving the plan is hard. The article argues that four kinds of state must be kept separately: run state, the current plan, in-flight tool calls, and learned memory. In-memory storage in LangGraph, Google ADK, and the OpenAI Agents SDK loses state at restart, while each framework documents a persistent alternative. Operations with unknown outcomes should not be retried blindly, and retries are only safe when the called service enforces an idempotency key. The author notes that the core difficulty lies in representing each state explicitly rather than in the absence of a database.