💭 in your head · nothing to type
Picture that session where Claude's answer was completely off — or it "forgot" a detail you'd already given.
Don't frame this as a system failure — picture it as a colleague back from holiday who remembers nothing of the project. No agreements, no context, no sense of why things are the way they are. Present, willing to work — but starting from a blank slate.
Run that session in your head: what exactly went on holiday along with their understanding? What would you need to explain to that colleague first?
💭 in your head · nothing to type
Agents don't have memory the way humans do. Everything they know right now is whatever sits in the current context window.
Don't picture this as a "memory failure" — picture it as a theatre prompter whispering from the wings. They hear only the line you slip them in this moment. Everything spoken in earlier acts is already beyond their earshot.
Hold that image: if your past messages were lines from other scenes — what would the prompter have missed tonight? What do you end up delivering as if for the very first time, every time?
Module 04 already covered these three failures and their fixes — see “The sins of prompting”. What is new here is not a repeat of the recipes but the agent's field of view.
Context isn't everything you've ever told the agent. It's what's in its window right now: this session's messages, plus the files and prompt you've given it.
Bad: 200 files in context + 2 hours of history
Good: 3–5 relevant files + a clean prompt for the task
💡 At any one moment the agent knows only what's in its "field of view". The way you manage that field is the way it works.