SafetyPack MCP v0.2.0 — standalone local memory

SafetyPack MCP 0.2.0

ローカルMCPのみを独立して公開しました。Alibaba/Qwen版のバックエンド、Web UI、クラウド設定は含みません。このリポジトリのソースZIPまたは添付wheelだけで導入できます。

• Windowsユーザーに紐づく暗号化ローカルメモリ。Copilot / Claude Code / Codexで同じ保存先を利用。
• 明示的な保存・検索・引き継ぎのみ。監視、常時収集、自動モデルダウンロードなし。
• 固定したローカル埋め込みモデル＋別モデルによる関連度再判定。索引作成は明示実行。
• 出典、提案と観測の区別、旧決定の明示的な置き換え、根拠不足・判定不一致の扱いを追加。

Install

Download the source ZIP below or clone this repository, then follow README.md. Alternatively, install safetypack_local-0.2.0-py3-none-any.whl in a Python 3.11+ virtual environment and run safetypack client-config. Windows is the default supported platform for DPAPI-backed storage. No Alibaba/Qwen account or API key is required.

Optional neural inference dependencies use the wheel [neural] extra (or .[neural] from source). Models are separate public downloads, approximately 259 MiB total: explicitly run safetypack neural-setup --allow-download, then safetypack reindex --project "C:\Projects\Example". Models and user memories are not bundled in this release.

Upgrade / limitations

Back up existing memory before writing with v0.2, then restart every MCP client on the new installation. Explicit writes upgrade the reader-format gate; old v0.1 clients fail closed. Memory keys and record IDs are preserved. Never expose recovery passphrases in chat or command arguments.

Search returns historical evidence, not verified answers. A relevant record may still omit the requested detail; answer_support remains not_assessed. Report missing details as not recorded. Small synthetic evaluations and their remaining misses/false retrievals are documented in README.md and tests/fixtures; there is no zero-error guarantee.

The older Alibaba/Qwen application remains at https://github.com/monoqlo78/safe-memory-platform and has been restored to its pre-MCP source tree, without rewriting history.