cm-sync

Keeps claude-mem memory the same on every computer, through a private git repo.

Each device publishes only the rows it created (as JSON Lines) and its work-view ledger in <memory repo>/claude-mem/devices/<device-id>/, and imports the other devices' rows through its own claude-mem worker (POST /api/import), so search and vector indexes stay correct. The one exception is work state (see below); nothing else writes to claude-mem.db directly. Python 3.9+ standard library only; macOS, Linux and Windows.

Install

git clone https://github.com/snuffish/cm-sync ~/cm-sync
ln -sf ~/cm-sync/cm-sync ~/.local/bin/cm-sync   # or any directory on PATH

On Windows, clone the same way and add the clone's folder to your user PATH; cm-sync.cmd runs it from cmd and PowerShell (through py -3, or python when the launcher is missing). Git Bash can use the cm-sync script as on macOS and Linux.

The memory repo defaults to ~/.claude-memory. If it doesn't exist yet, cm-sync setup asks for its git URL and clones it there. Override the path with CM_SYNC_REPO / --repo, and preset the URL with CM_SYNC_REMOTE (needed for setup --yes or non-interactive runs).

On a new machine:

cm-sync setup   # checks bun/node/git, the plugin and worker, clones the memory repo, turns on redaction, loads My Rules, names the device
cm-sync         # same as `cm-sync sync`

Commands

Command	Does
cm-sync / cm-sync sync	Pull → import → export → commit only this device's folder → push (sync is the default command)
cm-sync import	Pull the memory repo, then import other devices' rows and apply their deletions (--no-pull to skip the pull, --full after a database reset)
cm-sync export	Write this device's rows and work-view ledger to claude-mem/devices/<device-id>/ (--with-prompts to include raw prompts)
cm-sync status	Worker, redaction, active My Rules, work-view status (pending rule decisions, progress groups), counts per device, last push
cm-sync url	Print the worker URL (CLAUDE_MEM_WORKER_HOST/PORT from env, then settings, then 127.0.0.1:37700 + uid % 100, 37777 on Windows)
cm-sync hooks	Print Claude Code hooks that import on session start and sync on session end
cm-sync refile <session-id> <project>	Move a session's observations and summaries to another project (claude-mem files a session under the folder it started in)

How it works

• Work state (the lists behind claude-mem's "My Work" page) syncs too. No worker route keeps its timestamps, and claude-mem decides each task's current status by the latest timestamp, so cm-sync inserts other devices' entries into work_state_entries directly, the way claude-mem's own cloud sync does: original created_at_epoch, plus origin_device_id/origin_local_id so re-imports are no-ops and imported entries are never published again. It needs claude-mem v15+, and is skipped with a warning on older versions. The tool calls that link a list to the request it came from don't sync, so on other devices the lists can show under "work records need a linked human request" rather than as cards. Pick "All projects" and a wide enough time range to see older entries.
• My Rules sync with work state: rules you set, approve, edit, turn off or dismiss are work-state entries (list my-rules). claude-mem rebuilds ~/.claude-mem/MY-RULES.md only for rules its own cloud sync applies, so after every import cm-sync asks the worker to rebuild it (POST /api/work-view/rules/load-in-claude-code, claude-mem 15.1+). It does that only once ~/.claude/CLAUDE.md loads the file; cm-sync setup offers to add that line, so run it once on each device.
• Work-view ledger and rule suggestions sync too: ~/.claude-mem/work-view/ (story.json, request-groups.json, progress/, alignment.json) is published under each device's folder. On import, unreviewed rule suggestions ("Needs your decision") are shared across devices, and any rule resolved on any device (on, off, dismissed, replaced in work_state_entries) is automatically pruned from story.json. Request groups and conversations are additively unioned, progress entries pick the newest check-in, and alignment picks the latest.
• Deleting a row on the device that created it publishes a tombstone, and the other devices delete their copy on the next import.
• Export refuses to run while claude-mem redaction is off and skips sensitive observations. When a row looks like a secret it asks whether to delete it, otherwise it aborts (--exclude-flagged keeps those rows local instead).
• Only this device's folder is staged and committed, so other files in the memory repo are never touched.
• Local state (device id, imported/published/tombstoned keys) lives in ~/.local/state/cm-sync/ (%USERPROFILE%\.local\state\cm-sync on Windows).
• Files are written with LF line endings on every platform, so a Windows device's export is byte-identical to a Mac's.

If a secret reaches a commit anyway, deleting the row is not enough: rewrite the memory repo's history and re-clone it on every device.

git -C ~/.claude-memory filter-repo --path claude-mem/devices/<device-id>/<file> --invert-paths
git -C ~/.claude-memory push --force

Tests

python3 -I -m unittest discover -s ~/Projects/cm-sync -p 'test_*.py'   # `py -3` instead of python3 on Windows