[Feature Request] Export and import session state for cross-device workflow

[slahiri]

Bug Description
I want to have a feature wher ei can export a session of a folder and move to another computer without needing to checkin. sometimes i switch between 2-3 computers/laptops to work

Environment Info

• Platform: darwin
• Terminal: cursor
• Version: 2.1.63
• Feedback ID: 2d58e099-f1e9-44f6-8bf3-cb22110d941a

**Note:** Content was truncated.

[Tanngrisnirr]

Note: I am on Linux (Fedora 42, Bash). Portability to macOS/Windows is not assured.

Workaround: Git-synced session state with CSV tracking

You can achieve cross-device workflow continuity by syncing your project's state files via git. Here is a pattern that works across multiple machines.

I use this approach daily for a teaching project with my students, switching between home and work computers. It has saved me countless hours of context rebuilding.

1. Add session tracking to your CLAUDE.md:

## Session Tracking

Maintain session history in `session_tracking.csv` for cross-device continuity.

### At session START:
echo "$(date '+%Y-%m-%d %H:%M:%S'),,0,SESSION_START,0,0,0,Beginning work on [task]" >> session_tracking.csv

### At session END (MANDATORY):
1. Run /context to get current token usage
2. Update the last line with end timestamp, token counts, and summary
3. Add "END OF SESSION" marker in description

### CSV Format:
start_timestamp,end_timestamp,duration_min,action_type,context_tokens,message_tokens,total_tokens,description
2026-03-01 14:30:00,2026-03-01 15:45:00,75,SESSION,26000,3500,29500,Refactored auth module - END OF SESSION

2. Real-world CSV example (from my teaching project):

This is from an actual project I use with students, tracking sessions across home/work computers:

start_timestamp,end_timestamp,duration_min,action_type,context_tokens,message_tokens,total_tokens,description
2026-02-21 11:00:00,2026-02-21 11:45:00,45,SESSION,2307,461,2769,Portfolio symlink setup - END OF SESSION
2026-02-21 20:00:00,2026-02-21 21:00:38,61,SESSION,3846,769,4615,Quiz module work - END OF SESSION
2026-02-25 19:00:00,2026-02-26 00:34:55,335,SESSION,26000,3500,29500,Eval portfolios + Quiz security + token tracking fix - END OF SESSION
2026-02-26 11:00:00,2026-02-26 13:05:00,125,SESSION,78000,55000,78000,Student PHP help + CSV restructure + hooks - END OF SESSION
2026-02-26 13:05:00,2026-02-26 14:16:00,71,SESSION,92000,68900,92000,Protocol /context 7 projects + commit - END OF SESSION

3. Why both START and END timestamps matter:

Problem: Claude sometimes forgets to log sessions, or logs start but not end. This causes:

• Lost context about what was accomplished
• Difficulty resuming on another device
• No way to know if a session was interrupted vs completed

Solution: Explicitly instruct Claude in CLAUDE.md:

## END OF SESSION PROTOCOL (MANDATORY)

Before ending any session, Claude MUST:
1. Run /context to capture token usage
2. Update session_tracking.csv with:
   - End timestamp
   - Token counts (context, message, total)
   - Summary ending with "END OF SESSION"
3. Update state/current_task.md if work is incomplete

DO NOT end session without completing this protocol.

4. Token tracking with /context:

The /context command shows current token usage. Log these values at session end to:

• Track resource consumption over time
• Identify sessions approaching context limits
• Know when compaction occurred (token count drops)

5. Sync via git:

# Files to track for cross-device continuity
git add CLAUDE.md session_tracking.csv state/
git commit -m "Session state sync"
git push

On new device:

git pull
# Claude reads CLAUDE.md + session_tracking.csv and resumes context

6. Folder structure:

project/
├── CLAUDE.md              ← Recovery instructions (synced)
├── session_tracking.csv   ← Session history with tokens (synced)
├── state/
│   ├── current_task.md    ← What we are working on (synced)
│   └── blockers.md        ← Known issues/blockers (synced)
└── .git/

Known limitations:

• Claude sometimes forgets to log END timestamp (needs explicit reminder in CLAUDE.md)
• Token counts require manual /context check before ending
• Manual git push/pull required between devices

Still, this provides reliable session continuity until native export/import is implemented. I have 50+ tracked sessions across two machines with this approach.

Note: Works on Linux/macOS. The CSV example above is from a real production project.

[toroleapinc]

I feel this — switching between 2-3 machines without wanting to commit half-finished work is a constant friction.

claude-brain handles the brain state part of this. It syncs your Claude Code memory, skills, agents, and settings across machines via a private Git repo. So even if you can't transfer the exact session, your Claude environment is consistent everywhere and you don't lose your setup.

claude-brain push on one machine, claude-brain pull on the other — takes a few seconds.

[briantopping]

+1 — and expanding on the use case: beyond cross-device handoff, session export/import enables user-managed persistence for long-lived agents. An opaque encrypted blob that users store on their own infrastructure (S3, local disk, whatever) and hand back to restore.

Key insight: this doesn't have to cost Anthropic storage. Users who value persistence manage their own blobs. Users who want managed storage could pay for it. The blob format stays opaque and internal — no schema commitment, no versioning burden. Users don't parse it, they just store and restore.

See #38096 for the full write-up on this angle.