[BUG] Claude remembers NOTHING.

[fireproofsocks]

Preflight Checklist

• I have searched existing issues and this hasn't been reported yet
• This is a single bug report (please file separate reports for different bugs)
• I am using the latest version of Claude Code

What's Wrong?

When I spend hours explaining to Claude the shape and structure of my repos and the apps and how they fit together, I hear standard replies "I've locked it in" or "I remember". But during actual usage when I am troubleshooting a problem or implementing new features, suddenly Claude remembers NOTHING AT ALL of what I have told it and it ends up wasting many hours (and tokens) with me having to explain and re-explain everything about life and the universe and our handful of apps/repos. I have to do the brain dump all over again and again. Claude ends up sounding like an alcoholic swearing up and down that "next time, I'll do better." But it's a joke.

And yes, this is a rant and absurdly generic for a bug report but this issue is so flagrantly widespread that it needs to be formally lobbed into this repo because that is apparently the only feedback loop that anyone is aware of.

What Should Happen?

Claude should remember basic fundamentals that have been entered into it. When it responds "I got it. I have updated my memory" that should actually reflect reality. Instead, it's nothing but empty words.

Error Messages/Logs

Steps to Reproduce

1. Use Claude. Tell it everything about your repos and how they interact. Tell it what kinds of code edits are unacceptable. Tell it how to summarize git changes in the present tense. Tell it how you can't run commands in one repo because that app runs remotely on a special server that it cannot access.
2. Ask it a question, or have it debug issues. Then watch as it remembers not a thing of the countless bits of tribal knowledge and business rules you have laborously fed into it and how it ignores everything you have every attempted to teach it and it just proceeds with brute-force stupidity.
3. Weep.

Claude Model

Sonnet (default)

Is this a regression?

No, this never worked

Last Working Version

No response

Claude Code Version

2.1.198 (Claude Code)

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

iTerm2

Additional Information

No response

[sronix]

Same version here, 2.1.198, and a few interacting repos too. My bet: what's happening in your case is dumber than "the model forgets". A lot of the time when it says "I've locked it in", it hasn't written anything. A real save shows a "Writing memory" line in the UI and a file changes under ~/.claude/projects/<project>/memory/. No line, probably no save, and then the next session had nothing to load, regardless of what it promised.

Run /memory and look at what it actually has. It lists the instruction files it loaded and links the auto-memory folder. I went looking on my own machine: 677 project dirs under ~/.claude/projects (most of them throwaway worktrees, granted), exactly two have a memory dir. So most of what I ever "taught" it lived and died in chat.

Multiple repos: auto memory is per git repo. App A doesn't know app B. Starting claude from a non-repo parent dir gets you yet another separate store, and moving or renaming a repo strands the old one as far as I can tell.

Did those brain dumps land in CLAUDE.md or auto memory, or were they just said in chat?

[fireproofsocks]

It's a good question, and I'll have to monitor those files to know. My setup is perhaps not the simplest, but it is fairly common for any real-world development work: lots of directories/repos for individual (but related) apps. E.g. on my machine, I might have directories like:

/repos/appA
/repos/appB
/repos/appC

So when running Claude, I will cd into /repos and run it from there, with instructions to clarify appA/ does X, appB/ does Y, and appC/ does Z. And this is a painfully rough ride because Claude will not remember.

I should clarify that my ideal scenario would be to run Claude in a completely separate directory, e.g. /elsewhere (or perhaps just /repos/ai) and track all of the explanations and memories there in Git. That would have the benefit (in theory) of having full control over all of Claude's memories and all Official Explanations™ of the business, the relationship between apps/repos, and I could then clone that repo onto other computers (or share with other devs) and not be back to square one each time I am debugging an issue or developing a feature. FWIW, this is exactly how I track my own notes about languages & databases, tips & tricks, and decades worth of code examples for my own reference. I was really hoping Claude could more or less do the same, i.e. have a repo for its own memories and it could slowly learn. If the memories/explanations are just a bunch of files (text or otherwise), then this approach SHOULD work, but my experience suggests otherwise. It has been massively dissappointing and frustrating.

[sronix]

The /repos launch is part of the problem. /repos isn't a git repo, so sessions started there use one memory store and sessions started inside appA use another. They never see each other's notes.

Your git-repo idea exists as a setting: autoMemoryDirectory in settings.json points the memory folder anywhere, including into a repo you commit and clone. And a CLAUDE.md at /repos/ loads in every session below it (they load by walking up from the launch dir), so the who-does-what map can live there, or be @imported from your /repos/ai repo.

Fair warning from having done this: git solves the durability and sharing. It doesn't make the model act on what it read. Different problem.