Sync

Context Sync — Cross-Device + Cross-Tool

Two problems, one canonical source of truth:

1. Cross-device — Claude Code identifies projects by a hash of the absolute path, so context is lost between machines (/home/alice/myapp → -home-alice-myapp ≠ /Users/john/myapp → -Users-john-myapp). Solved by export / import.

1. Cross-tool — You use Claude Code, Cursor, Codex CLI, Roo, Cline, Aider, or Windsurf — each has its own instruction/memory file (CLAUDE.md, .cursorrules, AGENTS.md, ...). You re-explain your project to every one. Solved by tools push — one canonical memory file → adapter files for every detected tool.

The shared truth: .claude-context/MEMORY.md (+ memory/*.md), committed to git. Every tool reads from a locally-generated adapter file pointing at that truth. No SaaS. No proprietary memory broker. No vendor lock-in.

---

Commands

/sync export

Export current project context to .claude-context/ and commit.

Run:

python scripts/sync-export.py

What it does:

1. Calculates the Claude path hash for the current directory
2. Finds ~/.claude/projects// — reads all .md context files
3. Copies them to .claude-context/ in the project
4. Creates .claude-context/sync-manifest.json with device metadata
5. Runs git add .claude-context/ && git commit -m "context-sync: export "
6. Prints: push reminder

Then push:

git push

What gets exported:

• MEMORY.md — your project memory index
• memory/*.md — individual memory files
• project_*.md, feedback_*.md, user_*.md, reference_*.md — all context docs

What stays private (never exported):

• *.jsonl — conversation history (too large, private)
• *.json — may contain API keys or settings

---

/sync import

Restore project context on this device from .claude-context/.

Run:

python scripts/sync-import.py

What it does:

1. Runs git pull to get latest .claude-context/
2. Reads the manifest to confirm what was exported
3. Calculates the path hash for this device
4. Creates ~/.claude/projects// if needed
5. Copies all files from .claude-context/ to the correct local target
6. Prints a summary of restored files

After import: restart Claude Code or open the project. Claude immediately has full context.

---

/sync status

Show what has changed between the exported context and the current local context.

Check manually:

# What's in the local Claude folder
ls -la ~/.claude/projects/$(python3 -c "import os; print(os.path.abspath('.').replace('/', '-').replace('\\\\', '-').replace(':', '-'))")/

# What's been exported
ls -la .claude-context/

# Diff a specific file
diff .claude-context/MEMORY.md ~/.claude/projects/$(python3 -c "import os; print(os.path.abspath('.').replace('/', '-').replace('\\\\', '-').replace(':', '-'))")/MEMORY.md

Or ask Claude directly to compare them and summarize what's new, changed, or missing.

---

/sync clean

Remove large files from the local Claude project folder to reduce size before export.

Run:

# Show what's in your Claude project folder
python3 -c "
import os
from pathlib import Path
key = os.path.abspath('.').replace('/', '-').replace('\\\\', '-').replace(':', '-')
folder = Path.home() / '.claude' / 'projects' / key
if folder.exists():
    for f in sorted(folder.iterdir()):
        print(f'{f.stat().st_size:>10,}  {f.name}')
else:
    print('No Claude project folder found')
"

# Remove conversation history (*.jsonl) to free space
python3 -c "
import os
from pathlib import Path
key = os.path.abspath('.').replace('/', '-').replace('\\\\', '-').replace(':', '-')
folder = Path.home() / '.claude' / 'projects' / key
removed = 0
for f in folder.glob('*.jsonl'):
    size = f.stat().st_size
    f.unlink()
    print(f'Removed {f.name} ({size:,} bytes)')
    removed += size
print(f'Freed {removed:,} bytes')
"

---

/sync tools detect

List which AI dev tools are present in this project — shows what tools push would write.

Run:

python scripts/sync-tools.py detect

Detection (no false positives — we only write where the tool already lives):

| Tool | Output adapter file | Detected by | |---|---|---| | Claude Code | CLAUDE.md | always written (universal) | | Cursor | .cursor/rules/memory.mdc | .cursor/ or .cursorrules exists | | Roo Code | .roo/rules/memory.md | .roo/ or .rooignore exists | | Cline | .clinerules/memory.md | .clinerules exists | | Codex CLI | AGENTS.md | .codex/ or AGENTS.md exists | | Aider | CONVENTIONS.md | .aider/, CONVENTIONS.md, or .aider.conf.yml | | Windsurf | .windsurfrules | .windsurfrules or .windsurf/ exists |

---

/sync tools push

Generate adapter files from .claude-context/MEMORY.md for every detected tool.

Run:

python scripts/sync-tools.py push            # write them
python scripts/sync-tools.py push --dry-run  # show diff without writing
python scripts/sync-tools.py push --only cursor,claude  # subset

What it does:

1. Assembles the canonical memory (MEMORY.md + every memory/*.md) into one block
2. For each detected tool, wraps it in the right format (e.g. Cursor .mdc gets YAML frontmatter) with a clear "AUTO-GENERATED" header pointing back to the source
3. Writes .claude-context/sync-tools-manifest.json listing what was written

Why this is better than commercial cross-AI memory tools (Vilix, Mem0):

• No SaaS subscription
• Memory lives in your git repo, not a third-party DB
• One canonical file you can git blame, diff, and review like code
• New devs run python scripts/sync-tools.py push after git clone — every AI tool they use immediately has the project context
• When you update .claude-context/MEMORY.md, re-run push and every adapter regenerates

Why most adapter files are gitignored: They're per-machine projections of the canonical source. Committing them would create merge conflicts every time someone with a different tool stack syncs. The source-of-truth (.claude-context/) is what's committed; the adapters are local rebuilds.

Phase 2 — live MCP sync (planned, contributions welcome): See [PHASE2-MCP.md](PHASE2-MCP.md) for the full spec. Adds a small MCP server alongside the file adapters so edits flow between tools without re-running push. Covers every MCP-capable tool (Claude Code, Cursor, Codex CLI, Windsurf, Roo, Cline). The file-based Phase 1 stays as the universal fallback and the source of truth.

---

How the Path Hash Works

Claude Code generates the project key from the absolute path:

abs_path = os.path.abspath('.')
# /home/alice/myapp
# → replace '/' with '-'
# → -home-alice-myapp

claude_key = abs_path.replace('/', '-').replace('\\', '-').replace(':', '-')
source = f"~/.claude/projects/{claude_key}/"

| Device | Path | Hash | |--------|------|------| | Linux VPS | /home/alice/myapp | -home-alice-myapp | | Mac | /Users/john/myapp | -Users-john-myapp | | Windows | C:\Users\john\myapp | -C--Users-john-myapp |

These are all the same project but Claude sees three different contexts. sync-import.py handles this automatically — it always installs to the correct hash for the current device, regardless of where it was exported from.

---

Full Cross-Device Workflow

Device A (VPS):
  1. Open project in Claude Code
  2. Work — Claude builds up memory, project context
  3. /sync export → git push

Device B (Mac / laptop):
  1. git clone  (or git pull)
  2. /sync import
  3. Open project in Claude Code
  4. Claude remembers everything from Device A

Keep in sync:
  Device A: /sync export → git push   (after significant sessions)
  Device B: /sync import              (after git pull)

---

Setup (one-time)

The .claude-context/ folder is already part of the repo (committed). The scripts are in scripts/. No dependencies — pure Python stdlib.

Make sure .claude-context/ is not in .gitignore (this is intentional — it's the sync mechanism). The *.jsonl exclusion in .gitignore prevents conversation history from being committed.

---

What Gets Synced

| File type | Synced | Why | |-----------|--------|-----| | MEMORY.md | ✅ | Project memory index | | memory/*.md | ✅ | Individual memory entries | | project_*.md | ✅ | Project context docs | | feedback_*.md | ✅ | Behavioral feedback | | user_*.md | ✅ | User profile context | | reference_*.md | ✅ | Reference links | | *.jsonl | ❌ | Conversation history — private, too large | | *.json | ❌ | Settings — may contain keys |

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: veekunth217
• Source: veekunth217/claude-scaffold-skill
• License: MIT
• Homepage: https://github.com/veekunth217/claude-scaffold-skill

Install and usage instructions live in the source repository linked above.