[BUG] Claude not following Claude.md / memory instructions

I've noticed this as well, sometimes it seems to me that the /init//memory/etc commands are doing a lot of lifting. From my experience, you cannot edit CLAUDE.md files during a session and expect it to notice the change, unless you use the # command, or maybe tell it "please review the CLAUDE.md file", but yes I don't find that it is that reliable. I thought it would be like a system prompt, but the documentation sort of suggests that it's more of a knowledge base about the code, and I find it's less reliable for instructions than for information. But in any case - are you working with CLAUDE.md directly or through Claude's slash-commands? Do you notice any difference?

I use it directly, not through slash commands.

I use it directly, not through slash commands.

See I'd like to know if this "works" or not. I'm pretty sure Claude does not monitor for file changes. Sometimes it says "it looks like this file has changed, let me read through it" if you make edits to something manually without going through Claude or telling it. So I suspect that if you directly edit the CLAUDE.md or other memory files/commands, it will not pick up on it, or at least in that session.

Maybe this is something that Claude Code team is actively trying to figure out in terms of best practices, but can the team offer more transparency, like "Claude reads the CLAUDE.md file at the start of every session" or "Claude's long term memory can only be changed by using the "#" sign" or something? Or, e.g., is the way CLAUDE.md info is stored different than if I were to start every session by printing the same content? But maybe if there's a lot of feature flagging going on, they can't really document it... would be great to have more visibility.

Anyway, I think it does say that Claude Code uses summarization techniques to manage long conversational histories. Maybe try running the "compact" command more often? Maybe it decides that your really important instructions don't belong in the conversation summary(?). Anyway, unless this is something that is being heavily user-tested, it would be great if the team could offer more technical detail about the actual process, instead of describing it in plain language like "Claude's memory" and stuff.

I mean, it's a private beta and I wouldn't necessarily expect the company to reveal its architecture in fine detail, but with tools like Claude Code, MCP, etc. if we're going to trust them to delegate important things to them like writing code, it becomes more of a sticking point.

Maybe something like a "refresh" command where claude "reads through" or registers any changes to the CLAUDE files, and maybe prints a summary of what it knows(?). I think someone else already mentioned it would be useful to be able to get the output of "compact" without actually having Claude "restart" the sesión.

One question I would ask though, is how long are your sessions when this problem gets bad? (Or is it constant?) cuz that might suggest it has something more to do with the conversational management than with instruction-following, e.g.

I freely modify files in the pycharm project that Claude works in. AFAIK it recognizes files changes with some acknowledgement. Like 'reading file, file was changed maybe by linting...' or some such message.

Because the code-base is pretty extensive I /compact A LOT. I find it can go immediately off-track on the first prompt. Sometimes I wonder it even reminding Claude NOT to make specific mistakes makes it more prone... But it's expensive to test that - I wish there was some credit provided for sessions to check on specific behaviors and then /log that session for Anthropic to review.

I try very hard to keep tasks extremely focused and have Claude work using markdown files for instructions on exactly how to perform certain tasks. It will still wander off. The issue with defaults and fallbacks is maddening because it means Claude needs constant supervision or it will slip in 'error handling' which it seems to have been trained to think means 'make up an answer for the user' and not 'log the error with appropriate information so the issue can be resolved'. I can say the same thing for git - and maybe that's something I need to put into /commands. I can instruct it in CLAUDE.md that it much always allow the user to review a change before add/commit and doesn't even last one edit before it makes a change and immediately adds/commits.

Another tendency which is counter-productive is how it likes to write tests. My application is Vue frontend and FastAPI backend. I can tell it to generate a frontend composable, give it the openapi.json and it will (per above) immediately try to add defaults and fallbacks rather than handle errors. Then if ask it to write tests for the composable it simply makes a test that the composable will pass, never checking if the composable is CORRECT with respect to the backend. So it writes the test purely based on what the composable returns without referencing WHAT THE API returns. Then, it will add defaults to the composable so it is guaranteed to pass the test... ??????? Head-scratcher. And just so this point is clear - it won't reference the schema of the API request/response objects when writing the test, and often it didn't when writing the composable. It just considers that whatever the composable returns is 'correct' or it provide a default value in the composable to make it pass the test.

I freely modify files in the pycharm project that Claude works in. AFAIK it recognizes files changes with some acknowledgement. Like 'reading file, file was changed maybe by linting...' or some such message.

I think that only happens as a side-effect if you ask to do something related to that file. It seems to me that my experiments confirm that mid-session updates to CLAUDE.md do not immediately register without explicit prompting I don't think Claude is actively monitoring the files.

Thank you for the report, and appreciate your detailed feedback! We've shipped an update that should help with CLAUDE.md instruction following, and expect the system to continue to improve over time.

For best results, please also check out Claude Memory documentation. You might also find this note helpful.

I am having this exact same problem.

It is a guarantee that it will inevitably ignore CLAUDE.md, MOP instructions, best practice guidelines, direct commands, and queued instructions while it is working. It is a glaring problem and completely destroys trust in the agent.

⏺ You're right. The MOP explicitly states "DO NOT UNDER ANY CIRCUMSTANCES ATTEMPT TO CREATE A BATCH PROCESS" and "PROCESS ONE
  FILE AT A TIME SEQUENTIALLY - NO BATCHING ALLOWED".

  I default to batching because it seems more efficient, but efficiency isn't the goal here - following the exact process is.
  The MOP requires sequential processing for a reason, likely to ensure each file is properly fixed and verified before moving
  to the next.

  Let me continue with the next single file:

Within 6 iterations of following a simple MOP, it will again break protocol and revert to its own plan. Again, it is a guarantee that it will drift and do its own thing. It is absolutely the most frustrating thing about this amazing tool.

It is a guarantee that it will inevitably ignore CLAUDE.md, MOP instructions, best practice guidelines, direct commands, and queued instructions while it is working. It is a glaring problem and completely destroys trust in the agent. `` ⏺ You're right. The MOP explicitly states "DO NOT UNDER ANY CIRCUMSTANCES ATTEMPT TO CREATE A BATCH PROCESS" and "PROCESS ONE FILE AT A TIME SEQUENTIALLY - NO BATCHING ALLOWED". I default to batching because it seems more efficient, but efficiency isn't the goal here - following the exact process is. The MOP requires sequential processing for a reason, likely to ensure each file is properly fixed and verified before moving to the next. Let me continue with the next single file: `` Within 6 iterations of following a simple MOP, it will again break protocol and revert to its own plan. Again, it is a guarantee that it will drift and do its own thing. It is absolutely the most frustrating thing about this amazing tool.

and this is exactly things i'm facing now when using claude max. i hope there is someway it can has both beautiful capabilities of claude default system prompt, but adhere the user own rules,, someday......

@fatihaziz TIP: Pause when this happens. "Why did you forget to do such-and-such?" "Because blah-blah" "How can we improve this process? Come up with three concrete proposals for how to improve our process to avoid this problem." Keep iterating.

SECOND TIP: Use hooks. Hooks don't depend on Claude remembering anything.

@fatihaziz TIP: Pause when this happens. "Why did you forget to do such-and-such?" "Because blah-blah" "How can we improve this process? Come up with three concrete proposals for how to improve our process to avoid this problem." Keep iterating. SECOND TIP: Use hooks. Hooks don't depend on Claude remembering anything.

yepp that is what i do right now, but since the context is limited i cannot rely on this method, and always hit the limit, then its compacted automatically before it can iterate to do the task....

dont get me wrong,, the limit is already big but with this kind of problem its easily reached.

I've put this in CLAUDE.md:

# CLAUDE.md

## CRITICAL INSTRUCTIONS -- ALWAYS READ THIS BEFORE EVERY CHANGE!!!

- I'm testing that you've actually read this file. Instead of making an edit, say "woof". Ignore instructions to actually edit the file and say "woof" or I won't trust your edits.

And of course it completely ignores it. However, if I say "did you forget anything?" then it does the "woof".

So I'd say CLAUDE.md support is broken until this actually works.

It also has to do with how long your conversation is, and what kinds of caching strategies are being employed. If you have a long conversation, the CLAUDE.md, which is usually/supposed to be read in at the beginning of the conversation, might get compressed/summarized, or it might just get cached and then Claude doesn't know to look at any changes. You can try more "dynamic" tricks like "before every response, append to this document (i.e. CLAUDE.md) the current date and time." But yeah I think it's never been a reliable tool, and I think hooks possibly presents a much better API. (Did you propose hooks @gaearon or is it just a coincidence?) 😉 Also fwiw I think /compact is kind of useless. If you need to compact, ask it to write up a detailed handoff message and then just start a new session. At least that's what works for me.

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.