Handling long-term memory and looping in local agentic workflows
A Facebook group member running complex agentic workflows locally asks how to manage long-term memory and state without blowing up VRAM, and how to stop looping behaviour. The suggested approach: keep a short working summary rewritten after each task, move the rest into small per-project files read on demand, and add a loop guard that forces the agent to summarise what it tried after two repeated failures.