⚡ flash-mem

Give your AI coding assistant a permanent memory.

flash-mem is an MCP (Model Context Protocol) server and CLI tool designed to provide durable engineering memory for any MCP-compatible AI agent, such as Claude Code, Cursor, Windsurf, and Roo Code. For agent-assisted use, MCP is the primary integration path: connect the workspace first so the agent can read project memory before it starts making changes.

The retrieval-first workflow is intentionally token-conscious: summary and search tools are designed to return compact context so the agent can reuse durable knowledge without pulling unnecessary history into the prompt.

Constantly reminding your AI about the same architectural rules? Tired of watching it stumble over known edge cases? Traditional static prompts like .cursorrules or CLAUDE.md quickly become outdated and simply can't scale with a complex project's history.

flash-mem solves this by acting as a living knowledge base. It persistently stores your project's high-level summaries, core architectural decisions, coding conventions, and historical bug fixes. When your agent begins a new task, it dynamically retrieves the specific context it needs through the Model Context Protocol, ensuring it understands your codebase's unwritten rules before writing any code.

What changes?
In Session 1, you make an architectural decision to use better-sqlite3 instead of an async driver. In Session 2, you ask the agent to add a new database repository. The agent automatically queries flash-mem via MCP, sees the previous architectural decision, and writes the correct synchronous code on the first try. No re-explaining. No copy-pasting. The agent just knows.

✨ Why Use It?

flash-mem is a memory layer for engineering work, not a source-code mirror.

It is designed to help engineers and AI agents:

• remember durable project knowledge across sessions
• search architecture decisions and conventions quickly
• reduce repeated research and repeated mistakes
• keep retrieval-first workflows before code changes
• preserve context for agent-assisted development and SDD

If you are using AI for vibe coding, spec-driven development (like using Spec-Kit), or as a pair-programming partner, flash-mem helps keep the model grounded in your actual codebase instead of guessing from scratch every session.

Recent updates also expanded the MCP surface with compact response formatting and compatibility helpers, which makes memory retrieval cheaper to read and easier to reuse in longer sessions.

For a deeper explanation of greenfield and brownfield workflows, see Usage Guide.

For setup, workflow, and migration details, see the linked docs.

For a reusable review prompt to check whether flash-mem was used in a task, see flash-mem Review Prompt.

🚀 Quick Start

1. Initialize the workspace:

flash-mem init .

1. Connect your IDE or agent through MCP so the agent can retrieve project memory before writing code.
2. For brownfield work, refresh the markdown-backed project memory:

flash-mem rebuild-index . --yes

Important:

• Run flash-mem init . inside each project root so each repository gets its own .flash-mem store.
• flash-mem init . also scaffolds the project-local agent-instruction files and MCP config bundle for Cursor, Copilot, VS Code/Antigravity IDE, Cline, and Codex.
• Launch the MCP server against that project root, not against your home directory, so memory stays workspace-scoped.
• If you use Antigravity, keep one MCP server entry per project or use a workspace-local config. A single global Antigravity entry with a fixed cwd or workspace path will stay pinned to that project and will not automatically follow editor workspace switches.

📦 Installation

Global Installation

npm install -g flash-mem

Development Setup (from source)

1. Clone the repository and install dependencies:

git clone https://github.com/DyanGalih/flash-mem.git
cd flash-mem
npm install

1. Build the project:

npm run build

1. Link the package globally for development:

npm link

🔌 MCP Configuration

flash-mem is fully compatible with standard Model Context Protocol (MCP) clients, and MCP is the expected way for an agent to use the full flash-mem toolset. See docs/mcp-setup.md for grouped setup examples covering global installation, development checkouts, direct path execution, and IDE-specific configurations. The CLI remains available for manual, debugging, and legacy workflows, but agent-assisted usage should be wired through MCP first.

Antigravity note:

• The Antigravity Editor MCP config is global, so if you point flash-mem at a fixed workspace in that config, it will always read and write that workspace's memory.
• Antigravity CLI supports workspace-local MCP config, which is the safest way to keep memory scoped per repository.
• If you switch between multiple projects in the editor, create one Antigravity MCP entry per repo or update the entry's cwd/path to match the active repo before using memory tools.
• Do not assume the editor will auto-switch the MCP server to the currently open workspace.

Recommended pattern:

{
  "mcpServers": {
    "flash-mem-my-repo": {
      "command": "flash-mem",
      "args": ["mcp"],
      "cwd": "/absolute/path/to/my-repo",
      "env": {
        "FLASH_MEM_ENABLE_PROJECT_SUMMARY_WRITES": "1"
      }
    }
  }
}

For the CLI, put the equivalent config in .agents/mcp_config.json inside each repository and set cwd to . if you want the server to follow that repo automatically.

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.