# iai-mcp — Open-Source Memory System for AI Coding

> iai-mcp is an open-source local memory layer for AI coding assistants. It runs an MCP server plus capture/recall hooks and ships benchmarks.

## Install

Merge the JSON below into your `.mcp.json`:

## Quick Use

```bash
git clone https://github.com/CodeAbra/iai-mcp.git
cd iai-mcp
bash scripts/install.sh
export PATH="$HOME/.local/bin:$PATH"
iai-mcp capture-hooks install --target codex  # or: claude-code / all
iai-mcp doctor
```

## Intro

iai-mcp is an open-source local memory layer for AI coding assistants. It runs an MCP server plus capture/recall hooks and ships benchmarks.

**Best for:** Mac developers who want ambient memory with benchmarks and hooks

**Works with:** macOS, Python 3.11/3.12, Node.js 18+, Claude Code or Codex CLI

**Setup time:** 10-25 minutes

### Key facts (verified)

- GitHub: 116 stars · 12 forks · pushed 2026-05-13.
- License: MIT · owner avatar + repo URL verified via GitHub API.
- README-verified entrypoint: `git clone https://github.com/CodeAbra/iai-mcp.git && cd iai-mcp && bash scripts/install.sh`.

## Main

- Install the daemon + wrapper first, then install hooks; without hooks you get read access but not ambient capture/recall.

- Treat `iai-mcp doctor` + `daemon status` as your safety gates before enabling it in a real project session.

- If you care about numbers, run the repo’s benchmark modules to measure recall latency and fidelity on your machine.

### Source-backed notes

- README Install section uses `bash scripts/install.sh` and documents adding `~/.local/bin` to PATH for `iai-mcp`.
- README documents capture/recall hooks (`iai-mcp capture-hooks install|status|uninstall`) and shows a Codex target option.
- README includes benchmark claims (e.g., p95 <100ms, >=99% verbatim recall at 10k) and points to `bench/` modules.

### FAQ

- **Is it cross-platform?**: README lists macOS as supported and calls out Windows/Linux as not supported yet.
- **Does it store secrets?**: It stores your captured transcript locally; treat the store directory like any sensitive developer artifact.
- **What if the daemon is down?**: README describes hooks as fail-safe; session start should not block if recall returns empty.

## Source & Thanks

> Source: https://github.com/CodeAbra/iai-mcp
> License: MIT
> GitHub stars: 116 · forks: 12

---

<!-- ZH -->

## Quick Use

```bash
git clone https://github.com/CodeAbra/iai-mcp.git
cd iai-mcp
bash scripts/install.sh
export PATH="$HOME/.local/bin:$PATH"
iai-mcp capture-hooks install --target codex  # or: claude-code / all
iai-mcp doctor
```

## Intro

iai-mcp 是开源本地记忆层：作为 MCP server + 捕获/召回 hooks，为 Claude Code/Codex 等把历史对话变成可检索记忆。README 标注静态加密 AES-256-GCM、p95 召回 <100ms，并附带可复现基准。

**Best for:** 想要“自动记录+自动召回”，并且希望有基准数据支撑的 macOS 用户

**Works with:** macOS；Python 3.11/3.12；Node.js 18+；Claude Code 或 Codex CLI

**Setup time:** 10-25 minutes

### Key facts (verified)

- GitHub：116 stars · 12 forks；最近更新 2026-05-13。
- 许可证：MIT；作者头像与仓库链接均已通过 GitHub API 复核。
- README 中核对过的入口命令：`git clone https://github.com/CodeAbra/iai-mcp.git && cd iai-mcp && bash scripts/install.sh`。

## Main

- 先装 daemon + wrapper，再装 hooks；没有 hooks 只能读，无法做到“自动捕获/自动召回”。

- 把 `iai-mcp doctor` 与 `daemon status` 当作启用前的安全门槛，先确认环境与进程都正常。

- 如果你关心效果，用仓库自带的 benchmark 脚本在本机复测召回延迟与保真度。

### Source-backed notes

- README 的 Install 用 `bash scripts/install.sh`，并说明把 `~/.local/bin` 加进 PATH 后运行 `iai-mcp`。
- README 说明 capture/recall hooks（install/status/uninstall），并提供 Codex 作为 target 的参数示例。
- README 给出基准指标（如 p95 <100ms、10k 规模 verbatim recall ≥99%）并提供 `bench/` 可复现脚本。

### FAQ

- **跨平台吗？**：README 标注主要支持 macOS；Windows/Linux 暂未支持（以 README 为准）。
- **会存储敏感信息吗？**：它会本地保存对话捕获内容；请把存储目录当作敏感开发资料来管理。
- **daemon 挂了会怎样？**：README 描述 hooks 为 fail-safe：召回失败应返回空内容，不阻塞会话启动。

## Source & Thanks

> Source: https://github.com/CodeAbra/iai-mcp
> License: MIT
> GitHub stars: 116 · forks: 12


---
Source: https://tokrepo.com/en/workflows/iai-mcp-open-source-memory-system-for-ai-coding
Author: MCP Hub