memport 0.1.0

Make your AI coding tool's local auto-memory travel with the repo, via git.

memport

Your AI coding tool learns your project across sessions — then traps that knowledge on one developer's disk. memport sets it free, so it travels with the repo via git.

Claude Code accumulates auto-memory as it works: verified build commands, architecture notes, known issues, conventions. It lives at ~/.claude/projects/<encoded-git-root>/memory/ as clean markdown — machine-local, never synced. Clone a colleague's repo and you inherit CLAUDE.md and git history, but none of the tacit knowledge their Claude built up. memport closes that gap. It's a git subcommand.

Status: proof-of-concept, Claude Code only. Output is tool-agnostic; Cursor/Codex readers can be added later. Name is a placeholder (rename via the APP_NAME constant in config.py).

What it does (and doesn't)

Claude Code's memory files are already excellent structured markdown, so memport preserves them faithfully rather than distilling them into a lossy schema. The pipeline is:

~/.claude/.../memory/*.md   (per dev, per machine, never synced)
      │  git memport pull
      ▼
   strip frontmatter (drops local session ids)
   scrub (redacts secrets / real home paths; relative paths untouched)
      ▼
 .memport/memory/*.md   ← clean, committed, faithful
 .memport/review/*.md   ← anything scrub flagged, awaiting human sign-off
      │  git memport build
      ▼
   MEMPORT.md   ← @-imported by CLAUDE.md / AGENTS.md via marker block

Everything committed rides git's normal review/merge flow. Git is the sync layer — no server, no infra.

Install & use

pip install memport            # puts `git-memport` on PATH → `git memport`
# no-packaging path: chmod +x bin/git-memport && add bin/ to PATH
# or run directly: PYTHONPATH=src python3 -m memport <cmd>

git memport init               # scaffold .memport/, hook, wire adapters
git memport pull               # preserve + scrub local auto-memory
git memport status             # list anything flagged for review
git memport build              # regenerate MEMPORT.md (idempotent)
git add .memport MEMPORT.md CLAUDE.md && git commit

A post-merge hook re-runs build so teammates get updated memory on pull.

Design decisions

• Preserve, don't distill — real memory files are already well-formed.
• Derived index, never committed — no merge-conflict magnet.
• Own file (MEMPORT.md), imported via markers — never clobbers CLAUDE.md.
• Scrub is mandatory; findings route to review/ for a human gate.
• Frontmatter stripped — drops machine-local session ids.

Related

Anthropic issue #28276 requests native in-repo memory storage. If shipped it gives memport a clean redirect target; the scrub / cross-tool / curation value is unaffected.

License

Apache-2.0