Long-Context Agent Development: The agent might miss the requirements or the detail , and loss systematical view of the structure

Hi there:
   Do you meet the problem that the agent(claude code with minimax-v2.1). After 3-5 compacts, the model totally loss the memory and do not have systematic view about the whole system, and start to write shit, Which is really annoyed, Do you have any suggestions ? 
