I quit the infinite session — why I now close Claude every few hours of work

2026-06-11 ／ Vol 15 ／ draft at the time of publishing

Let me start with my own words. Asked how I use Claude these days, this is what I say:

I use Claude by feel, testing as I go. There was a time I tried to run a single session forever, but now I close the session after every few hours of work. I think this is better. No problems with how it feels to use, either.

Today's post is about how I landed on "close it every few hours." It's a short story, but if you work with an AI every day, you probably reach the same fork in the road.

At first, it was the infinite session

When I started with Claude Code, I tried to stretch one session as long as I possibly could. Monthly bookkeeping, site fixes, random lookups — everything went into the same continuing conversation. It wasn't so much a deliberate choice as the fact that I couldn't find a reason to close it. If anything, a long-running session felt like an asset I had grown — letting it go felt like a waste.

Reasons not to close it, I had plenty. Closing a session means losing the context. "The approach we just settled on." "The circumstances I already explained." "The unspoken agreement about how we handle this kind of thing." As long as the conversation keeps going, Claude remembers all of it. Close it, and it's gone.

Why closing felt scary

The fear, in one sentence: I hated the idea of re-explaining everything from zero, every time. That was the whole of it.

My work is full of preconditions. Running an EC shop, three websites, the monthly books. Each one carries dozens of accumulated "here, we do it this way" decisions. Reciting all of that at the start of every new session sounded exhausting just to imagine. So I didn't close. And because I didn't close, everything got stuffed into one conversation. It grew longer and messier, and still I couldn't let go — "because the context is in there." Looking back, I think I was treating the session itself like a notebook I had been filling up.

External memory arrived, and the fear went away

The turning point was that the machinery for keeping context outside the conversation came together. Three pieces, concretely — none of them clever tricks, just decisions about where things get written down and what gets read first:

1) Claude's persistent memory. Per-project memory lives in files and gets loaded at the start of every new session. Thanks to the multi-PC memory sharing setup I wrote about in Vol 9, every machine starts from the same memory.

2) A CLAUDE.md in each repository. The working conventions for each site and project — policies, the do-not-do list, procedures, past history — written on a single page. The rule is that a new session reads it before touching anything.

3) The habit of leaving build logs and runbooks as files. Whenever something gets built, the how and the why get written to a file on the spot. The answer to "wait, how did we do that last time?" lives in a file, not in a conversation log. The next time the same work comes around, that file is the script.

With these three in place, the context lives on disk, not inside the conversation. Closing a session no longer destroys it. A new session reads the memory and the CLAUDE.md, catches up in seconds, and picks up almost exactly where the last one left off. At some point I noticed the fear of closing was simply gone.

The side effect: closing forces checkpoints and records

There's one more effect I only noticed by doing it. The act of closing a session forces the work to have a boundary and a record.

Before closing, "here's what got done, here's what's next" has to be put into words and saved — to memory, to a build log, somewhere. Skip that and the next session suffers, so it naturally becomes a habit. The result: unfinished tasks no longer lie scattered and forgotten inside an endless conversation. Putting the checkpoint into words doubles as a handover note — to the next session, and to my future self.

The infinite-session era was the opposite. The conversation never ended, so the work never ended either. "How far did we get?" could only be answered by scrolling back through the conversation — and even the scrolling-back was delegated to Claude. Work without boundaries doesn't get recorded. I only saw this after I started closing.

I still don't know if this is the right answer

To be honest, I'm not convinced this is the answer. Some work probably suits one long session better — a continuous deep-dive into a single problem dies if you cut it in the middle. On days like that, I still let the session run long.

All I can say today is that "close at every few hours of work" fits the way I currently use it. No problems with the feel of it. If that changes, the change itself becomes a post.