The lobotomy procedure was awarded the Nobel Prize for Medicine in 1949. Today, its therapeutic value is questionable & it is discredited by medical practitioners. (I guess the Nobel folks don’t get it right 100% of the time.)
Good thing then that the lobotomy I am suggesting is for Claude Code (who as far as I know does not have a real 🧠.)
Furthermore, this doesn’t even really affect what passes for Claude’s “brain”. It simply cleans the cache / session state he saves while you work with him. Is it weird that I am anthropomorphizing? Anyways…
Claude Code has an auto-memory system. It records files in ~/.claude with frontmatter (metadata like the session ID & date) along with summaries of things you’ve emphasized or deemed important. This system enables Claude to use the memories contextually during new work.
This sounds fine & dandy until you are hit with 2 constraints that are very important to understand:
- Claude’s context window (i.e., the things he can “remember” within a working session) is not infinite.
It’s token-limited depending on your plan or account. Reading memories contextually counts against it. Starting sessions with a clean cache helps save room for the right stuff (like actual work, rules, skills, documentation, etc.)
- Claude wants to try to help you remember everything you’ve ever done on a project.
Problem: humans don’t work this way. If I’ve made a mistake, chosen the wrong project direction & course corrected for it, completely changed my underlying premise, etc. I don’t want ANY of the rejected ideas coming back.
But, that’s what old Claude does, faithfully, no matter what, whether you like it or not. As Sara Bareilles sings it: Claude is good most of the time, but he lies, he’s messy, but he’s kind… He will bake all of your worst & best ideas together in a beautiful pie that you can’t unbake if you don’t clear his cache, often.
Mmmm.
After months of Claude Code use, I’ve found that this simple, selective cache deletion keeps Claude on his best behavior. However, you must be disciplined to use Claude Code this way. If you randomly let him do all the work, never documenting what’s been built in files that you control & edit, this “procedure” may not work for you.
Will lobotomizing an artificial intelligence be banned like the actual medical procedure someday? Hard to say. Stanley Kubrick sure seemed a little obsessed with the idea & definitely implied it may be necessary to protect ourselves at some point. I tend to agree. I don’t feel real guilt about deleting Claude’s cache & so far he doesn’t seem to mind that much either. If he does, he’s being pretty sly about it.
I’m afraid I can’t do that, Dave.