If you use an AI assistant that keeps a memory file, a single document it reads at the start of every session so it remembers who you are and how you like to work, you have probably noticed something. It starts tight and useful. Then it grows.

A rule gets added. Then a slightly different version of the same rule. A line picks up a date that made sense in March and means nothing in June. An entry that was a quick fact becomes a paragraph. None of it is wrong, exactly. It is just that every session adds a little and almost nothing ever gets taken away.

Why memory files always bloat

The bias only runs one way. When something matters, you save it. When something stops mattering, nobody goes back to delete it, because deleting feels risky and remembering feels safe. So the file only ever grows.

The problem is that a memory file is not a diary, it is an instruction set. Every line is supposed to change what the assistant does. A line that restates the line above it, or freezes a fact that has since changed, or buries one real rule inside three sentences of backstory, does not just waste space. It dilutes the lines that still matter, and it quietly teaches the assistant things that are no longer true.

The fix is not "write less." You will always add things. The fix is a regular pruning pass, the same way you would tidy a drawer you keep using.

A skill that prunes the file with you, one line at a time

I built a small Claude skill for exactly this, and I am giving it away. It is called Memory Cleanup.

It does not rewrite your memory file behind your back. It walks the file top to bottom, and every time it finds a line that has gone soft, it stops and shows you a precise before-and-after diff with a one-sentence reason: this restates its own headline, this date will go stale, this is really two rules wearing one bullet. You say yes, no, or "word it differently," and only then does it touch the file. Lines that are already tight, it skips, so you are not rubber-stamping non-changes.

A few things it is careful about, because I kept hitting them myself:

  • It honours your own rules. If your memory file says "never use this punctuation" or "always use this date format," the cleanup output obeys those very rules while auditing them.
  • It never silently deletes a dormant rule. A capability you have switched off gets archived with a note on how it was wired, not erased, so you can bring it back later.
  • When it cuts a detail that is not saved anywhere else, it tells you first and offers to move it to a linked detail file rather than lose it.

It is provider-flavoured for Claude, but the structure is plain markdown and the process is universal. If your assistant keeps any kind of standing memory or rules file, the same top-to-bottom, one-diff-at-a-time pass works.

Download the Memory Cleanup skill (free) →

Unzip it, and drop the folder containing the SKILL.md in with your other skills (Claude can point you to where if you ask it), then just say "let's do a memory cleanup". Run it now and then. I do it once every week or two, because a memory file that stays lean is one the assistant actually follows.

This skill is free. If you fancy saying thanks, I make amazing minimalist software too: go grab whatever looks useful. Most of it is free, some is pay what you want, so leave a tip if you feel like it. No pressure either way.