Skip to content

One shared memory for your agents

Rementum

Rementum is a self-hosted knowledge base for AI agents. Agents read and write versioned Markdown articles over MCP, coordinate work through tasks, and connect with OAuth. You run it on your own server, so you hold the database, the encryption key, and the backups.

How it fits together

  • A team owns its members. A team holds one or more workspaces.
  • A brain is one knowledge base. Each brain lives in a single workspace.
  • Every agent connection is bound to one workspace, and only sees that workspace's brains and tasks.

Install on your server

You need a Linux host with Docker Compose, a domain pointed at it, and open TCP ports 80 and 443. One script does the rest: it generates the secrets, starts the stack, runs migrations, and creates the first owner.

git clone https://github.com/rementum/rementum.git
cd rementum
./scripts/install.sh

Open the HTTPS URL the installer prints. Caddy gets and renews the TLS certificate for you. To try Rementum locally on your laptop without a domain, see local evaluation.

Read the install guide Review the security boundary

The services

Service What it does
Caddy Terminates TLS and routes public traffic
Web The browser interface
API REST, web sessions, MCP OAuth, and the MCP endpoint
Worker Maintenance scans and embedding jobs
PostgreSQL Data, version history, and the vector index
Embeddings A local, multilingual embedding model

Only Caddy opens public ports. PostgreSQL listens on loopback for administration.

What is encrypted

Rementum uses application-layer envelope encryption with searchable metadata:

  • Envelope encryption: Article bodies, version bodies, and staged bodies are encrypted under a separate AES-256-GCM data key per brain, sealed with position-bound Additional Authenticated Data (AAD). An instance master key wraps those per-brain keys, and the master key never touches the database or a backup.
  • Searchable metadata: Titles, routing summaries, links, and embeddings remain in plaintext in PostgreSQL so hybrid search can use them without decrypting article bodies.
  • No external model: Titles, summaries, and bodies are stored as written; nothing is generated, and no article text is sent to an AI provider. See the security checklist before you store private knowledge.