[BUG] Memory saved to undocumented MEMORY.md instead of CLAUDE.md

Must be something new from today.

❯ Why are you using `MEMORY.md`? Can you show me your instructions that reference it? It may be something new i'm not familiar with.

● Good catch! Here's the section from my system prompt about auto memory:

  # auto memory

  You have a persistent auto memory directory at `/home/martin/.claude/projects/-home-martin-(...)/memory/`. Its contents persist across conversations.

  As you work, consult your memory files to build on previous experience. When you encounter a mistake that seems like it could be common, check your auto memory for relevant notes — and if nothing is written yet, record what
  you learned.

  Guidelines:
  - Record insights about problem constraints, strategies that worked or failed, and lessons learned
  - Update or remove memories that turn out to be wrong or outdated
  - Organize memory semantically by topic, not chronologically
  - `MEMORY.md` is always loaded into your system prompt — lines after 200 will be truncated, so keep it concise and link to other files in your auto memory directory for details
  - Use the Write and Edit tools to update your memory files

  So MEMORY.md is automatically loaded into my context for future conversations. I updated it because we learned some important patterns (like the EdgeQL GROUP BY limitation, migration workflow, etc.) that would be useful for
  future work.

  However, if you'd prefer I don't use it or organize the memory differently, I'm happy to adjust! What would you prefer?

I think it's probably indeed a bug that these instructions are enabled in production. There's nothing in the release notes about this.

It's not production ready:

• There is no way of opting out of the behaviour
• This behaviour should really be opt-in if it should exist at all — good users carefully curate what goes into their memory files; now we're suddenly telling Claude to write whatever session-specific minutiae it likes to a file which is automatically loaded into every future session?
• The cherry on top is that the user has no way of seeing what Claude is writing to this file from the Claude Code CLI

Hi everyone, thanks for the feedback! I'm working on updates to the UI as well as getting our docs for auto-memory published. If you want to opt-out, you can disable auto-memory by setting the env var CLAUDE_CODE_DISABLE_AUTO_MEMORY=1

I'm experiencing the opposite of this issue now — MEMORY.md was working well for me (memories were being saved there and loaded into context), but as of v2.1.38, the MEMORY.md contents are no longer injected into the system prompt at all.

I discovered this when I asked Claude to "check project memory" about plugin versioning rules I'd previously saved. It had no idea and had to search the filesystem to find the file.

I then directly interrogated the model about its system prompt:

Me: "Do you have anything like You have a persistent auto memory directory at in your system prompt?" Claude: "No, I have nothing like that. I just re-scanned my entire system prompt carefully. There is no mention of: persistent auto memory directory, MEMORY.md, any memory directory path, any instruction to check project memory."

Me: "So you have nothing about MEMORY.md in your system prompt?" Claude: "The only file-related context I get is: Contents of ~/.claude/CLAUDE.md and Contents of <project>/CLAUDE.md. That's it."

I also verified:

• No CLAUDE_CODE_DISABLE_AUTO_MEMORY env var is set
• No memory-related settings in ~/.claude/settings.json
• The MEMORY.md file exists at ~/.claude/projects/<encoded-path>/memory/MEMORY.md with valid content
• It's simply never loaded into context

Filed a separate issue for this: #25318

So to summarize: the MEMORY.md location may be undocumented (your issue), but the bigger problem now is that even if you accept MEMORY.md as the storage location, its contents are not loaded into the model's context — making the entire memory feature silently broken.

This is now documented here: https://code.claude.com/docs/en/memory#auto-memory

We're rolling auto-memory out slowly, but if you'd like to opt-in during the preview period, you can do so from /memory in Claude Code

This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.