Reducing Context Switching Cost with Running Notes - How AI Agents Solve the Same Problem
Reducing Context Switching Cost with Running Notes - How AI Agents Solve the Same Problem
The context switching cost is real, and it is mostly about losing your mental model. You are deep in a refactor, you understand the dependency chain, you know which three files need to change together. Then someone pings you about a bug. Twenty minutes later you come back to the refactor and you have lost it. You have to rebuild the mental model from scratch.
The Running Notes Fix
What helped was keeping a running notes file. Not a to-do list - a stream of consciousness document that captures your current mental model:
- What you are working on right now
- Why you made the last decision
- What you were about to do next
- Which files are involved and how they connect
When you get interrupted, you spend 30 seconds writing down where you are. When you come back, you read the last few lines and your mental model reconstructs in minutes instead of the usual fifteen to twenty minutes of re-exploration.
AI Agents Have the Same Problem
AI agents face the exact same challenge. Every new session starts with zero context. The agent does not know what it worked on yesterday, what decisions were made, or what constraints matter. It has to rebuild its mental model every single time.
The solution is the same - a running notes file. In the Claude Code ecosystem, this is CLAUDE.md. It stores decisions, preferences, constraints, and project context that persists across sessions. Instead of the agent rediscovering that "we use tabs not spaces" or "the auth module is being deprecated" every time, it reads the file and starts with context.
Why This Pattern Works
Both humans and AI agents share a fundamental limitation - working memory is finite and volatile. The fix in both cases is externalizing the mental model into a persistent document.
The difference is that humans usually resist writing things down because it feels like overhead. AI agents have no such resistance - if you put it in CLAUDE.md, they read it every time. This makes the AI version of running notes more reliable than the human version.
The best developer workflows combine both: a human running notes file for your own context, and a CLAUDE.md for your AI agent's context. Two persistent memories working in parallel.
Fazm is an open source macOS AI agent. Open source on GitHub.