Impulse: Claude Code’s Memory Isn’t a Feature, It’s a Burden

Claude Code’s memory isn’t a feature, it’s a burden.

If I could change one thing about Claude Code right now, it would be how it handles memories. I’m firmly convinced that the memory system makes Claude Code worse, not better.

A few days ago, Claude implemented an API of mine incorrectly — twice. I blocked the resulting tool calls with the note: “That’s not how the API works. Stop guessing.” (My patience was limited that day, hence the phrasing.)

In such cases, Claude Code nudges the model to “learn” from the rejection and its reason. So it dutifully writes down a memory. It’s a bit cute, honestly.

The trouble is, these memories are often either too specific or too general — and almost never on point. In my API case, Claude then explained to me in the following sessions that it had explicitly made sure to implement the API correctly — exactly as it had learned. Unfortunately, this new case was a different one, and the implementation was wrong again.

This happens regularly. “Memories” lull Claude into a false sense of security about its knowledge of the codebase. More of these memories means less exploration. They are feral snippets of documentation that go stale and may never have been correct in the first place. They take up tokens in the context that I, as an engineer, could put to better use.

(Incidentally, a million context tokens is no help here either. Anything above 80,000 tokens still isn’t good enough; memories or not.)

Not every addition is worthwhile. New features in Claude Code have to be configurable and remain predictable in the long run, to give experienced engineers the greatest possible leverage. I’ll bet many of you are good enough at efficiently providing the right context in every new session. Your Claude doesn’t need automatic memories, because you have them and you are in the driver’s seat.

PS: If you feel the same way I do and find the automatic memories too eager, you can manage or delete them via /memory and then “Open auto-memory folder”.

PPS: Even Milla Jovovich (yes, the actress) is building a solution to the “memory problem” of coding agents together with a colleague. I, on the other hand, don’t even see the missing memories as a problem in my daily work. And you?