Open Source · MIT License · Self-Hosted

Give your AI agent
a memory that lasts.

SecondBrain is a persistent MCP memory server. Your agents remember context across every session - stored in a semantic wiki, a knowledge graph, and a vector index you fully control.

Get started in minutes View on GitHub

agent session

# Your agent remembers across every session

agent › remember("The user prefers dark mode and uses Neovim")

✓ Stored. Task ID: a3f8c1

# Three sessions later…

agent › recall("user preferences")

↳ Dark mode enabled. Editor: Neovim. Prefers concise answers.

Why SecondBrain

Memory that actually scales

Built-in memory engines and simple lists work fine at first. Then they don't.

📝

Flat memory lists

Claude memory, ChatGPT memory

📄

LLM-Wiki / Notion

Simple wiki dumps

⬤

SecondBrain

HybridRAG + knowledge graph

Scales past 100+ facts

✗ Context fills up

~ Gets noisy

✓ Vector index + graph

Finds related concepts

✗ Keyword only

✗ Manual links

✓ Semantic + graph traversal

Resolves contradictions

✗ Keeps both

✗ No awareness

✓ Hourly review agent

Human-readable format

✗ Opaque store

✓ Markdown

✓ Obsidian-compatible vault

You own the data

✗ Vendor cloud

~ Depends on tool

✓ Fully self-hosted

Works with any agent

✗ Locked to one app

✗ Manual copy-paste

✓ MCP protocol

Features

Everything your agent needs to remember

SecondBrain is not a simple key-value store. It's a full knowledge management system designed for AI agents.

📈

HybridRAG Retrieval

When your agent calls recall(), SecondBrain runs three retrieval passes in parallel - vector similarity search (Qdrant), knowledge graph traversal (Neo4j), and direct Markdown lookup - then synthesizes a single coherent answer with an LLM.

Qdrant Neo4j LLM Synthesis

📄

Wikipedia-Agent Model

A planning agent decides how to update the knowledge base - creating, extending, or merging wiki pages intelligently rather than dumping raw text.

🔗

MCP-Native

Two tools: remember() and recall(). Drop into claude.ai, OpenClaw, or any MCP-compatible client with a single URL.

📚

Obsidian-Compatible Vault

All knowledge is stored as plain Markdown with [[wikilinks]]. Sync to any Git host. Browse it in Obsidian. Your data, your format.

⏰

Hourly Review Agent

A background agent automatically cross-links related pages, resolves contradictions, and keeps the knowledge base coherent over time.

🏠

Fully Self-Hosted

One docker compose up. Runs on your server. No third-party cloud, no data sharing, no subscription. MIT licensed.

Architecture

Simple API. Powerful internals.

Your agent sees two MCP tools. Under the hood, a multi-layer system handles everything.

Agent calls `remember(text)`

The input is split into independent topics and queued for async processing.

Planning agent decides

An LLM agent searches existing pages and determines what to create, update, or merge.

Three stores updated

Markdown vault, Neo4j knowledge graph, and Qdrant vector index are all updated atomically.

Agent calls `recall(query)`

HybridRAG searches all three layers and returns a synthesized, contextual answer.

Your Agent

claude.ai · OpenClaw · custom

↓ MCP

MCP Server

remember()

recall()

↓

Celery Worker

↓

📄 Vault

🔗 Neo4j

📈 Qdrant

↓

HybridRAG

↓

LLM Answer

Quick Start

Up and running in 3 steps

Docker Compose handles all dependencies - Neo4j, Qdrant, Redis, backend, and frontend.

Clone & configure

bash

git clone https://github.com/JonasKruegerCode/SecondBrain.git
cd SecondBrain
cp .env.example .env
# Set OPENROUTER_API_KEY and MCP_API_KEY in .env

Start

bash

docker compose up -d
# Web UI:  http://localhost
# MCP:     http://localhost:3000/mcp

Connect claude.ai

Settings → Integrations → Add MCP server

URL:     https://mcp.your-domain.com/mcp
API Key: your-secret-key

For local testing: expose port 3000 with ngrok or Cloudflare Tunnel.

Open Source

Help us make it production-ready

SecondBrain is MIT-licensed and actively seeking contributors. Here's where you can have the most impact:

🔧

Alternative LLM Providers

Support direct OpenAI / Anthropic keys without requiring OpenRouter as a proxy.

Enhancement

📄

Documentation & Cookbook

Usage examples and patterns for common agent workflows - the most impactful contribution for new users.

Docs

🔗

Git Sync: Any Host

GitLab, Gitea, and self-hosted support. Currently only GitHub PAT auth is tested.

Feature

🧪

Tests & CI

Expand integration test coverage and add GitHub Actions for pytest, ruff, and mypy on PRs.

Testing

Star on GitHub

Open an issue before starting work on a significant change.

Give your AI agent a memory that lasts.