Context management and working memory

[laoliu5280]

Agents lose track during long runs. They forget earlier instructions, drift off task, or even lose their working directory.

Tasks

• Maintain a STATE.md file in the workspace that the agent updates at each phase (current phase, what is done, key findings, next steps)
• Add periodic working directory checks (agents losing their directory is a known failure mode)
• Before each phase, generate a summary of prior phases to keep the context focused
• Validate that each phase produced its expected outputs before starting the next one
• Limit the agent to the top-K most promising directions instead of exploring everything

Research questions

• What information matters most across phases?
• How to detect context drift during execution?
• What is the right tradeoff between context breadth and depth?