A session is more than a chat history you can scroll later. It’s the active room where work happens. In Claude Code or Codex, it includes your instructions, the files the agent has read, the tool results it has seen, the mistakes it corrected, and the decisions you made along the way.
Think of a session like a working meeting. You sit down, put the relevant papers on the table, talk through the job, and produce work. If it runs six hours, covers seven unrelated topics, and nobody writes down what changed, the meeting gets messy. AI sessions work the same way.
How it shows up
Every session has a context window, so the AI doesn’t have infinite working memory. If you ask it to write copy, debug code, review a contract, plan a training, and process a transcript all at once, you’re asking one meeting to hold five. Our practical rule is to use sessions to hold phases of work: one for research, another for drafting. When a phase is done, close it cleanly and hand the next session the useful output instead of dragging the whole conversation forward. This is also why compaction matters: when a session fills up, the tool may summarize to keep going, but a summary isn’t the full meeting and details can get dropped.
Why you care
A good session feels focused. You tell Claude Code its role, what files matter, what outcome you want, and when to stop. At the end you either ship the output or turn the useful context into something durable through transcript processing, a note, or a skill. The quality of the meeting shapes the quality of the work.