内容
# git-mcp-memory
Git-backed long-term memory for AI agents, served through MCP and a JSON-first CLI.
<p align="center">
<img src="docs/assets/header.svg" alt="git-mcp-memory header illustration" width="100%">
</p>
<p align="center">
<a href="https://github.com/tomohiro-owada/gmem/actions/workflows/ci.yml"><img alt="CI" src="https://img.shields.io/github/actions/workflow/status/tomohiro-owada/gmem/ci.yml?branch=main&style=flat-square&label=ci"></a>
<a href="https://github.com/tomohiro-owada/gmem/releases/latest"><img alt="Release" src="https://img.shields.io/github/v/release/tomohiro-owada/gmem?style=flat-square"></a>
<img alt="Go" src="https://img.shields.io/badge/go-1.24%2B-00ADD8?style=flat-square&logo=go&logoColor=white">
<img alt="MCP" src="https://img.shields.io/badge/MCP-stdio-0f766e?style=flat-square">
<img alt="Embeddings" src="https://img.shields.io/badge/embeddings-local%20ONNX-2563eb?style=flat-square">
<img alt="Storage" src="https://img.shields.io/badge/source%20of%20truth-Git%20Markdown-f97316?style=flat-square">
</p>
<p align="center">
<a href="#installation">Installation</a>
·
<a href="#mcp-usage">MCP Usage</a>
·
<a href="#cli-usage">CLI Usage</a>
·
<a href="docs/design.md">Design Notes</a>
·
<a href="#日本語">日本語</a>
</p>
`git-mcp-memory` stores the source of truth as Markdown files in a Git repository, while using SQLite and local embeddings as a rebuildable search index. It is designed for local AI-agent workflows where memory should be readable by humans, reviewable with Git history, and searchable by meaning.
English | [日本語](#日本語)
## Features
- **Git as the source of truth**: every memory is an append-only Markdown file committed to a normal Git repository.
- **SQLite as a cache**: the local index can be deleted and rebuilt from Markdown.
- **Local vector search**: embeddings run in-process through ONNX Runtime.
- **No Ollama or embedding API server required**: model files are downloaded on first use and cached locally.
- **MCP over stdio**: works as a local MCP server for AI clients.
- **JSON-first CLI**: the same operations are available from the command line.
- **Project-aware search**: memories are grouped by project derived from the workspace path.
- **Cross-project search**: search all stored memories when needed.
- **Safety gate**: common secrets and personal information are rejected before saving.
- **Failure-aware Git behavior**: local commits are preserved when push fails, and `retry-push` can recover later.
## Status
This repository is usable but still young. The workflow is optimized for local personal use, but it is designed to tolerate multiple users or multiple machines sharing the same memory repository:
- one main branch in the memory repository
- append-only memory files
- `git pull --rebase` before writes and retry-push recovery after push failures
- local SQLite index
- local ONNX embedding model
- MCP stdio transport
See [docs/design.md](docs/design.md) for the design notes and tradeoffs.
## Architecture
```text
AI client / CLI
|
| save_memory / search_memory
v
git-mcp-memory
|
| Git pull / commit / push
| Markdown source files
| SQLite index
| ONNX embeddings
v
Memory Git repository
```
Memory repository layout:
```text
memory-repo/
└── projects/
└── my-project-a1b2c3d4/
└── decision-title_20260622_101530_a1b2c3.md
```
Each memory file uses Markdown with YAML front matter:
```markdown
---
type: Memory
title: "Decision title"
description: "Short description"
resource: null
tags: []
timestamp: 2026-06-22T10:15:30Z
project_id: "my-project-a1b2c3d4"
source: "mcp"
---
Markdown body...
```
## Requirements
- Go 1.24+
- Git
- Access to a Git remote repository for memory storage
- SSH or HTTPS authentication already configured for that remote
- Network access on first use to download:
- `intfloat/multilingual-e5-small` from Hugging Face
- ONNX Runtime from the Microsoft GitHub release
The embedding provider is currently fixed to:
- provider: `builtin_onnx`
- model: `intfloat/multilingual-e5-small`
- dimension: 384
## Installation
Download a prebuilt archive from the latest GitHub Release.
macOS arm64:
```bash
curl -L -O https://github.com/tomohiro-owada/gmem/releases/latest/download/git-mcp-memory-darwin-arm64.tar.gz
tar -xzf git-mcp-memory-darwin-arm64.tar.gz
mkdir -p ~/.local/bin
mv git-mcp-memory-darwin-arm64 ~/.local/bin/git-mcp-memory
chmod +x ~/.local/bin/git-mcp-memory
```
Linux amd64:
```bash
curl -L -O https://github.com/tomohiro-owada/gmem/releases/latest/download/git-mcp-memory-linux-amd64.tar.gz
tar -xzf git-mcp-memory-linux-amd64.tar.gz
mkdir -p ~/.local/bin
mv git-mcp-memory-linux-amd64 ~/.local/bin/git-mcp-memory
chmod +x ~/.local/bin/git-mcp-memory
```
Verify the install:
```bash
git-mcp-memory schema --output json
```
Release artifacts include a `.tar.gz` archive and a `.sha256` checksum file. Source builds are still supported:
```bash
go build -ldflags '-s -w' -o ~/.local/bin/git-mcp-memory ./cmd/git-mcp-memory
```
## Configuration
Create a JSON config file at the default location for your OS.
Default config path:
- macOS: `~/Library/Application Support/git-mcp-memory/config.json`
- Windows: `%LOCALAPPDATA%\git-mcp-memory\config.json`
- Linux: `${XDG_CONFIG_HOME:-~/.config}/git-mcp-memory/config.json`
Example:
```json
{
"git_dir": "/Users/alice/Library/Application Support/git-mcp-memory/repo",
"remote_url": "git@github.com:alice/my-memory-repo.git",
"index_path": "/Users/alice/Library/Application Support/git-mcp-memory/index.sqlite",
"embedding_provider": "builtin_onnx",
"embedding_model": "multilingual-e5-small",
"embedding_model_repo": "intfloat/multilingual-e5-small",
"embedding_model_revision": "main",
"embedding_query_prefix": "query: ",
"embedding_document_prefix": "passage: ",
"limits": {
"max_title_bytes": 512,
"max_content_bytes": 65536,
"hard_max_content_bytes": 1048576
},
"security_policy": {
"reject_personal_information": true,
"reject_organization_names": true,
"reject_customer_names": true
}
}
```
Notes:
- `remote_url` must point to an existing Git repository.
- Repository creation and SSH key management are intentionally outside this tool.
- If `git_dir` does not exist, the tool clones `remote_url`.
- SQLite is not committed to Git. It is a local cache.
## MCP Usage
Run the MCP server:
```bash
git-mcp-memory mcp
```
Example Codex-style MCP configuration:
```toml
[mcp_servers.gmem]
command = "/Users/alice/.local/bin/git-mcp-memory"
args = ["mcp"]
startup_timeout_sec = 120
```
Available tools:
### `save_memory`
Input:
```json
{
"current_workspace_path": "/path/to/project",
"title": "Decision title",
"content": "Markdown body",
"dry_run": false
}
```
Behavior:
1. Derives `project_id` from the workspace folder and Git remote/path hash.
2. Rejects unsafe content.
3. Generates a local embedding.
4. Pulls the memory repo with rebase.
5. Writes a new Markdown file.
6. Commits and pushes.
7. Updates the SQLite index after a successful push.
### `search_memory`
Input:
```json
{
"query": "What did we decide about local embeddings?",
"current_workspace_path": "/path/to/project",
"limit": 5,
"fields": ["title", "path", "content"],
"snippet_chars": 300
}
```
Use `"all": true` instead of `current_workspace_path` for cross-project search.
### `retry_push`
Retries pushing local commits that were preserved after a push failure.
```json
{
"dry_run": true
}
```
## CLI Usage
The CLI is designed for AI agents first. JSON output is the default.
Show schema:
```bash
git-mcp-memory schema --output json
```
Check status:
```bash
git-mcp-memory status --output json
```
On a fresh machine, `assets_ready` is usually `false` until the first embedding model setup completes. The first `save`, `search`, or `sync` command downloads the ONNX model, tokenizer, and ONNX Runtime into the local application data directory, so it can take longer than normal. Agents can check these fields before a real operation:
```json
{
"assets_ready": false,
"embedding_model_ready": false,
"tokenizer_ready": false,
"onnx_runtime_ready": false
```
After the files are cached, the same fields become `true` and later operations do not download them again.
Save a memory:
```bash
git-mcp-memory save \
--workspace /path/to/project \
--title "Decision title" \
--content "Markdown body" \
--output json
```
Dry run:
```bash
git-mcp-memory save \
--workspace /path/to/project \
--title "Dry run" \
--content "Validate and embed only." \
--dry-run \
--output json
```
Search within the current project:
```bash
git-mcp-memory search "local embeddings" \
--workspace /path/to/project \
--limit 5 \
--output json
```
Search all projects:
```bash
git-mcp-memory search "incident summary" --all --limit 10 --output json
```
Return one JSON object per result:
```bash
git-mcp-memory search "push failure" --all --output ndjson
```
Rebuild/synchronize local state:
```bash
git-mcp-memory sync --output json
```
Retry failed pushes:
```bash
git-mcp-memory retry-push --output json
```
## Agent Skill
This repository includes a Codex-style agent skill for CLI operation:
```text
agents/skills/git-mcp-memory-cli/
```
Use it when an agent needs to call `git-mcp-memory` through the CLI rather than through MCP tools. The skill covers JSON-first command usage, dry-run saves, project and cross-project search, status checks, and push-failure recovery.
## Safety Model
`save_memory` and `save` run a server-side security gate before writing anything.
Always rejected:
- private keys
- AWS access keys
- GitHub tokens
- OpenAI keys
- Slack tokens
- bearer tokens
- env-style secret assignments such as `TOKEN=...`
- invalid UTF-8
- unsafe control characters
Rejected by default, configurable by policy:
- email addresses
- phone numbers
- organization/customer labels detected by the built-in rules
Rejected values are not echoed back in responses. Responses include categories and fields, not the matched secret text.
## Git Failure Behavior
The tool optimizes for preserving memory without hiding synchronization problems.
- Save first creates a local Git commit.
- Push is attempted immediately.
- If push fails, the local commit remains.
- The response contains `pushed: false` and a structured `push_failed` warning.
- Search continues using local repository state when possible and returns a warning.
- `retry-push` can be used after the network or authentication is fixed.
Example warning:
```json
{
"code": "sync_failed_local_results",
"message": "retry push failed; search results are based on local repository state",
"details": {
"recommended_action": "retry_push",
"unpushed_commit_count": 1
}
}
```
## Development
Run tests:
```bash
go test ./...
```
### Build:
```bash
go build -ldflags '-s -w' -o ./bin/git-mcp-memory ./cmd/git-mcp-memory
```
Run a local smoke test:
```bash
git-mcp-memory save \
--workspace "$PWD" \
--title "Smoke test" \
--content "This is a local smoke test." \
--dry-run \
--output json
```
## Release
CI runs on pushes and pull requests to `main`.
Releases are created by pushing a `v*` tag:
```bash
git tag vX.Y.Z
git push origin vX.Y.Z
```
The release workflow builds native artifacts on GitHub-hosted Linux and macOS runners, uploads checksums, and publishes a GitHub Release.
##
English は [こちら](#git-mcp-memory).
### git-mcp-memory
AI agent の長期記憶を Git 管理された Markdown として保存し、MCP と JSON-first な CLI から読み書きするためのローカルツールです。
`git-mcp-memory` は、記憶の正本を Git リポジトリ内の Markdown に置き、SQLite とローカル embedding を再生成可能な検索インデックスとして使います。人間が読めること、Git の履歴で追えること、意味検索できることを同時に満たす設計です。
## 特徴
- **正本は Git**: 記憶は append-only な Markdown ファイルとして commit されます。
- **SQLite は cache**: 壊れたり削除したりしても Markdown から再構築できます。
- **ローカル vector search**: ONNX Runtime でプロセス内推論します。
- **Ollama や外部 embedding API server は不要**: 初回利用時にモデルをダウンロードして cache します。
- **MCP stdio 対応**: ローカル MCP server として AI client から使えます。
- **CLI も同じ機能を提供**: AI agent が叩きやすい JSON 出力を標準にしています。
- **project 単位の検索**: workspace path から `project_id` を導出します。
- **全 project 横断検索**: 必要に応じて全記憶を検索できます。
- **保存前 safety gate**: secret や個人情報を保存前に拒否します。
- **push 失敗に強い**: push に失敗しても local commit を残し、後から `retry-push` で復旧できます。
## 現在の位置づけ
この repository は利用可能ですが、まだ若い実装です。ローカル個人利用に最適化していますが、同じ memory repository を複数人または複数端末で共有する運用にも耐える設計です。
- memory repository は `main` branch を直線的に使う
- 記憶ファイルは append-only
- 書き込み前に `git pull --rebase` し、push 失敗時は `retry-push` で復旧する
- SQLite は local index
- embedding は local ONNX model
- MCP transport は stdio
詳細な設計メモは [docs/design.md](docs/design.md) を参照してください。
## 構成
```text
AI client / CLI
|
| save_memory / search_memory
v
git-mcp-memory
|
| Git pull / commit / push
| Markdown source files
| SQLite index
| ONNX embeddings
v
Memory Git repository
```
memory repository の例:
```text
memory-repo/
└── projects/
└── my-project-a1b2c3d4/
└── decision-title_20260622_101530_a1b2c3.md
```
各 memory file は YAML front matter 付き Markdown です。
## 必要なもの
- Go 1.24+
- Git
- memory 保存用の Git remote repository
- その remote へ push できる SSH または HTTPS 認証
- 初回利用時の model download 用 network access
現在の embedding model:
- provider: `builtin_onnx`
- model: `intfloat/multilingual-e5-small`
- dimension: 384
## インストール
最新の GitHub Release から build 済み archive を取得します。
macOS arm64:
```bash
curl -L -O https://github.com/tomohiro-owada/gmem/releases/latest/download/git-mcp-memory-darwin-arm64.tar.gz
tar -xzf git-mcp-memory-darwin-arm64.tar.gz
mkdir -p ~/.local/bin
mv git-mcp-memory-darwin-arm64 ~/.local/bin/git-mcp-memory
chmod +x ~/.local/bin/git-mcp-memory
```
Linux amd64:
```bash
curl -L -O https://github.com/tomohiro-owada/gmem/releases/latest/download/git-mcp-memory-linux-amd64.tar.gz
tar -xzf git-mcp-memory-linux-amd64.tar.gz
mkdir -p ~/.local/bin
mv git-mcp-memory-linux-amd64 ~/.local/bin/git-mcp-memory
chmod +x ~/.local/bin/git-mcp-memory
```
確認:
```bash
git-mcp-memory schema --output json
```
release artifact には `.tar.gz` archive と `.sha256` checksum を含めます。source からの build も可能です。
```bash
go build -ldflags '-s -w' -o ~/.local/bin/git-mcp-memory ./cmd/git-mcp-memory
```
## 設定
OS ごとの標準 config path:
- macOS: `~/Library/Application Support/git-mcp-memory/config.json`
- Windows: `%LOCALAPPDATA%\git-mcp-memory\config.json`
- Linux: `${XDG_CONFIG_HOME:-~/.config}/git-mcp-memory/config.json`
設定例:
```json
{
"git_dir": "/Users/alice/Library/Application Support/git-mcp-memory/repo",
"remote_url": "git@github.com:alice/my-memory-repo.git",
"index_path": "/Users/alice/Library/Application Support/git-mcp-memory/index.sqlite",
"embedding_provider": "builtin_onnx",
"embedding_model": "multilingual-e5-small",
"embedding_model_repo": "intfloat/multilingual-e5-small",
"embedding_model_revision": "main",
"embedding_query_prefix": "query: ",
"embedding_document_prefix": "passage: ",
"limits": {
"max_title_bytes": 512,
"max_content_bytes": 65536,
"hard_max_content_bytes": 1048576
},
"security_policy": {
"reject_personal_information": true,
"reject_organization_names": true,
"reject_customer_names": true
}
}
```
補足:
- `remote_url` は作成済みの Git repository を指定してください。
- GitHub repository 作成や SSH key 作成はこの tool では行いません。
- `git_dir` が存在しなければ `remote_url` から clone します。
- SQLite は Git に commit しません。local cache として扱います。
## MCP として使う
MCP server 起動:
```bash
git-mcp-memory mcp
```
MCP client 設定例:
```toml
[mcp_servers.gmem]
command = "/Users/alice/.local/bin/git-mcp-memory"
args = ["mcp"]
startup_timeout_sec = 120
```
利用できる tool:
- `save_memory`
- `search_memory`
- `retry_push`
`save_memory` は保存前に安全性検査、embedding 生成、Git pull、Markdown 作成、commit、push、index 更新を行います。
`search_memory` は Git の同期、SQLite 再 index、query embedding 生成、vector search を行います。
`retry_push` は push 失敗時に残った local commit の再送に使います。
## CLI として使う
schema:
```bash
git-mcp-memory schema --output json
```
status:
```bash
git-mcp-memory status --output json
```
初回利用前は、たいてい `assets_ready` が `false` です。最初の `save`、`search`、`sync` で ONNX model、tokenizer、ONNX Runtime を local application data directory へダウンロードするため、通常より時間がかかります。AI agent は本番操作の前に次の field を見れば、初回 setup 中かどうかを判断できます。
```json
{
"assets_ready": false,
"embedding_model_ready": false,
"tokenizer_ready": false,
"onnx_runtime_ready": false
}
```
cache 済みになるとこれらは `true` になり、以降の操作では再ダウンロードしません。
保存:
```bash
git-mcp-memory save \
--workspace /path/to/project \
--title "Decision title" \
--content "Markdown body" \
--output json
```
dry-run:
```bash
git-mcp-memory save \
--workspace /path/to/project \
--title "Dry run" \
--content "Validate and embed only." \
--dry-run \
--output json
```
project 検索:
```bash
git-mcp-memory search "local embeddings" \
--workspace /path/to/project \
--limit 5 \
--output json
```
全 project 横断検索:
```bash
git-mcp-memory search "incident summary" --all --limit 10 --output json
```
同期と再 index:
```bash
git-mcp-memory sync --output json
```
push 再試行:
```bash
git-mcp-memory retry-push --output json
```
## Agent Skill
この repository には CLI 操作用の Codex-style agent skill を同梱しています。
```text
agents/skills/git-mcp-memory-cli/
```
MCP tool ではなく CLI から `git-mcp-memory` を使う agent 向けです。JSON-first な command 利用、dry-run 保存、project 検索、全 project 横断検索、status 確認、push 失敗時の復旧手順を含みます。
## セキュリティ
`save_memory` と `save` は保存前に server-side の safety gate を通します。
常に拒否されるもの:
- private key
- AWS access key
- GitHub token
- OpenAI key
- Slack token
- bearer token
- `TOKEN=...` のような env secret
- invalid UTF-8
- 危険な control character
デフォルトで拒否されるもの:
- email address
- phone number
- built-in rule で検出される会社名、顧客名 label
拒否された値そのものは response に含めません。category と field だけを返します。
## Git 障害時の挙動
この tool は、同期失敗を隠さず、かつ local の記憶を失わないことを重視します。
- 保存時は local Git commit を作成します。
- push は即座に試みます。
- push に失敗しても local commit は残します。
- response には `pushed: false` と `push_failed` warning を返します。
- search は可能な限り local repository state で継続し、warning を返します。
- network や認証を直した後に `retry-push` で復旧できます。
## 開発
test:
```bash
go test ./...
```
build:
```bash
go build -o ./bin/git-mcp-memory ./cmd/git-mcp-memory
```
smoke test:
```bash
git-mcp-memory save \
--workspace "$PWD" \
--title "Smoke test" \
--content "This is a local smoke test." \
--dry-run \
--output json
```
## Release
CI は `main` への push と pull request で実行されます。
release は `v*` tag を push すると作成されます。
```bash
git tag vX.Y.Z
git push origin vX.Y.Z
```
release workflow は GitHub-hosted Linux/macOS runner で native artifact を build し、checksum と一緒に GitHub Release へ公開します。
MCP 配置
以下是此 MCP Server 的配置信息，可直接复制到 Cursor 或其他 MCP 客户端中使用。
mcp.json
连接信息
猜你喜欢
暂无推荐服务器