What is M-flow?

M-flow is not a vector database or a RAG framework. It is a cognitive memory engine — a system that doesn't just store text, but understands it, distilling incoming knowledge into structured layers and weaving them into the right place of a persistent architecture.

Where traditional retrieval matches queries by surface similarity, M-flow reasons through relationships — tracing connections, weighing context, and reaching answers the way cognition does.

Most AI memory systems fall into two categories:

• Context Window — The model re-reads everything from scratch. Exhaustive, linear, forgetful.
• Similarity Search — External memory exists, but retrieval is surface-level — matching by vector shape, blind to structure. Approximate, lossy.

M-flow introduces a third paradigm: a Cognitive Engine where memory becomes a mind. A query at any granularity finds a precise anchor, then expands outward — surfacing related context and associative content across different levels of detail.

How It Works

M-flow organizes knowledge into a four-level Cone Graph — a layered hierarchy from abstract summaries to atomic facts:

Retrieval is graph-routed: the system casts a wide net across all levels, projects hits into the knowledge graph, propagates cost along every possible path, and scores each Episode by its tightest chain of evidence. One strong path is enough — the way a single association triggers an entire memory.

For the full technical deep-dive, see Retrieval Architecture | 检索架构详解（中文）

Benchmarks

All benchmarks use gpt-5-mini (answer) + gpt-4o-mini (judge) + top-k=10.

LoCoMo-10

LongMemEval

Per-category breakdowns, reproduction scripts, raw data, and methodology for all systems: mflow-benchmarks

Features

Retrieval modes: Episodic is the primary retrieval mode — it uses graph-routed Bundle Search for best accuracy and is used in all benchmarks. Triplet Completion is a simpler vector-based mode suited for customization and secondary development. See Retrieval Architecture for details.

Quick Start

One-Command Setup (Docker)

git clone https://github.com/FlowElement-ai/m_flow.git && cd m_flow
./quickstart.sh

The script checks your environment, configures API keys interactively, and starts the full stack (backend + frontend). On Windows, use .\quickstart.ps1.

Install via pip

pip install mflow-ai         # or: uv pip install mflow-ai
export LLM_API_KEY="sk-..."

Install from Source

git clone https://github.com/FlowElement-ai/m_flow.git && cd m_flow
pip install -e .             # editable install for development

Run

import asyncio
import m_flow


async def main():
    await m_flow.add("M-flow builds persistent memory for AI agents.")
    await m_flow.memorize()

    results = await m_flow.search("How does M-flow work?")
    for r in results:
        print(r)


asyncio.run(main())

CLI

mflow add "M-flow builds persistent memory for AI agents."
mflow memorize
mflow search "How does M-flow work?"
mflow -ui          # Launch the local web console

Architecture Overview

┌───────────────┐     ┌───────────────┐     ┌───────────────┐
│  Data Input   │────▶│    Extract    │────▶│   Memorize    │
│  (50+ formats)│     │  (chunking,   │     │  (KG build,   │
│               │     │   parsing)    │     │  embeddings)  │
└───────────────┘     └───────────────┘     └───────┬───────┘
                                                    │
                      ┌───────────────┐     ┌───────▼───────┐
                      │    Search     │◀────│     Load      │
                      │  (graph-routed│     │   (graph +    │
                      │  bundle search│     │  vector DB)   │
                      └───────────────┘     └───────────────┘

Project Layout

m_flow/              Core Python library & API
├── api/             FastAPI routers (add, memorize, search, …)
├── cli/             Command-line interface (`mflow`)
├── adapters/        DB adapters (graph, vector, cache)
├── core/            Domain models (Episode, Facet, FacetPoint, …)
├── memory/          Memory processing (episodic, procedural)
├── retrieval/       Search & retrieval algorithms
├── pipeline/        Composable pipeline tasks & orchestration
├── auth/            Authentication & multi-tenancy
├── shared/          Logging, settings, cross-cutting utilities
└── tests/           Unit & integration tests

m_flow-frontend/     Next.js web console
m_flow-mcp/          Model Context Protocol server
mflow_workers/       Distributed execution helpers (Modal, workers)
examples/            Runnable example scripts
docs/                Architecture & design documents

Development

git clone https://github.com/FlowElement-ai/m_flow.git && cd m_flow
uv sync --dev --all-extras --reinstall

# Test
PYTHONPATH=. uv run pytest m_flow/tests/unit/ -v

# Lint
uv run ruff check . && uv run ruff format .

See CONTRIBUTING.md for the full contributor guide.

Deployment

Docker

docker compose up                       # Backend only
docker compose --profile ui up          # Backend + frontend
docker compose --profile neo4j up       # Backend + Neo4j

MCP Server

cd m_flow-mcp
uv sync --dev --all-extras
uv run python src/server.py --transport sse

Testing

PYTHONPATH=. pytest m_flow/tests/unit/ -v        # ~963 test cases
PYTHONPATH=. pytest m_flow/tests/integration/ -v  # Needs .env with API keys

Contributing

We welcome contributions! Please see CONTRIBUTING.md for guidelines, and our Code of Conduct for community standards.

License

M-flow is licensed under the Apache License 2.0.

Copyright 2026 Junting Hua

Licensed under the Apache License, Version 2.0.
You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0