Bug: Observation system creates CLAUDE.md files in project subdirectories and duplicate nested directories

[yungweng]

Description

The claude-mem plugin (v9.0.1) is creating CLAUDE.md observation files in project subdirectories where work is being done, and occasionally creates duplicate nested directories (e.g., backend/backend/, frontend/frontend/).

Steps to Reproduce

1. Work on files in a project subdirectory (e.g., backend/simulator/iot/)
2. Make edits to multiple files in that directory
3. Check git status after the session

Observed Behavior

Spurious files appear:

?? backend/backend/
?? backend/database/migrations/CLAUDE.md
?? backend/seed/fixed/CLAUDE.md
?? backend/simulator/iot/CLAUDE.md

Each generated CLAUDE.md file contains:

<claude-mem-context>
# Recent Activity

<!-- This section is auto-generated by claude-mem. Edit content outside the tags. -->

*No recent activity*
</claude-mem-context>

Or populated observation tables from the session.

Expected Behavior

• Observation data should be stored in the claude-mem database (~/.claude-mem/claude-mem.db), not written to project directories
• No files should be created in the user's project working directory
• No duplicate nested directories should be created

Environment

• claude-mem version: 9.0.1
• Platform: macOS (Darwin 25.2.0)
• Claude Code: Latest

Additional Context

The duplicate directory issue (e.g., backend/backend/) appears to be a path handling bug where the directory name gets doubled during some operation.

This behavior pollutes the working directory with untracked files and requires manual cleanup after each session.

Workaround

Manually delete the files after each session:

find . -path "./.git" -prune -o -name "CLAUDE.md" -newer <start-of-session> -type f -print
# Then selectively delete spurious ones

Or add to .gitignore (but this hides legitimate CLAUDE.md files too).

[niteeshm]

This is HIGHLY annoying and should have been handled as part of the initial release. I have added it to my .gitignore

[LukasAndreano]

OMG the SAME. I'm so tired of this.

[coletoncodes]

Experiencing the same issue. Quite frustrating.

[thedotmack]

This is intentional because even though it looks messy, it's actually a HYPER optimized token saving method that follows Anthropic's best practices to the T.

I'm adding right now the fixes for this though, to make sure it's part of a build step, optional setting actually was supposed to be default but I think it missed that

I know it's ugly, and I knew it would be an issue, but the benefit outweighs the messiness and I already see a huge gain in accuracy by leaving it as is

[coletoncodes]

I can see the benefits, I guess it just falls on pushing updates and ignoring this properly. Is the expectation to keep this per session and not longer term?

Impacts I see are on non-solo projects where some team members like myself are using this plugin and other are not. It bloats and confuses the team dynamic is all.

For solo projects it's absolutely incredible.

[niteeshm]

This is intentional because even though it looks messy, it's actually a HYPER optimized token saving method that follows Anthropic's best practices to the T.

Sure but this is at the cost of polluting the user's repo which is not the best way probably.

[itinance]

Have the same thing, commented it in this issue (i thought it is related, maybe it's not): #697 (comment)

[joefitter]

further to this, if I am working on 2 projects at once, claude.md files leak across and it creates irrelevant directories in the wrong project

[iefnaf]

@thedotmack Is there a switch to control this behavior?

[EYYCHEEV]

PR #823 should fix this - it adds a check for CLAUDE_MEM_FOLDER_CLAUDEMD_ENABLED which defaults to false, making folder CLAUDE.md generation opt-in.

+1 for getting this merged! 🙏