MemFlow

MemFlow finds relevant skills for your task in Claude Code. Register your SKILL.md files, then describe your task as usual. MemFlow searches your skills and gives Claude a short list of matches to read through MCP.

The current workflow uses Qdrant and an OpenAI-compatible embedding service. The default retrieval path needs no generation LLM. Optional LLM reranking uses a separate configured endpoint. MemFlow stores and serves the SKILL.md text; accompanying scripts and reference files are not included.

Quick start

You need Python 3.10+, uv 0.12.24, Claude Code, and Docker with Compose. You also need an existing OpenAI-compatible embedding service. The commands work in a Linux/macOS shell.

1. Install MemFlow

git clone https://github.com/skhynix/MemFlow.git
cd MemFlow
uv tool install .
docker compose -f scripts/docker-compose.qdrant.yml up -d

uv tool install makes memflow available from any project. If uv reports that its executable directory is missing from your PATH, run uv tool update-shell and open a new terminal. See uv’s tool installation guide.

Qdrant runs at http://localhost:6333 and keeps its data in scripts/qdrant_data/. If you already have Qdrant, skip the Docker command and use your server’s URL below.

2. Add connection settings to your project

Go to the project where you use Claude Code:

cd /path/to/your-project
vi .env

Add the following entries to .env, keeping any settings your project already has. The endpoint, model, dimensions, and API key below are example values; replace them with your embedding service’s settings.

QDRANT_BASE_URL=http://localhost:6333
QDRANT_COLLECTION_NAME=memflow_skills

VECTOR_EMBEDDING_API_BASE=https://your_endpoint/v1
VECTOR_EMBEDDING_MODEL=Qwen/Qwen3-Embedding-4B
VECTOR_EMBEDDING_DIMENSIONS=2560
VECTOR_EMBEDDING_API_KEY=EMPTY

Use EMPTY if your service does not require an API key. .env.example includes optional settings. Add .env to your project’s .gitignore to keep credentials out of Git.

3. Connect Claude Code

From your project’s root directory, run:

memflow claude configure --apply

This installs the hooks, registers a private MCP server for this project, and saves the path to your .env. Skill commands, the hook, and MCP reuse that file. You can safely run the command again with the same settings.

The output should show hook.installed_after: true and mcp.installed_after: true.

4. Register a skill

Use a skill you already have. The example below uses skills/release-checklist; replace it with your skill’s directory or SKILL.md path.

Create a file in your project:

mkdir -p skills/release-checklist
vi skills/release-checklist/SKILL.md