context: what the agent actually sees

the working memory

Everything the agent knows about your project right now lives in one place: the context window (the amount of text the model can consider at once – your messages, the files it read, the outputs of its tools). Not on disk, not in a database – in the conversation itself. What entered the context, the agent sees; what didn't, does not exist for it.

Four things fill it:

1. the standing rules – CLAUDE.md / AGENTS.md, read at session start (previous lesson)
2. your messages – every request, every correction
3. what the agent read – files, search results, documentation
4. what its tools returned – test output, build errors, screenshots