Rementum

One versioned, auditable brain behind every AI agent.

Self-hosted shared memory for Claude, Codex, Cursor, and any remote MCP client.

English · Türkçe · 中文

Documentation · Install · Security · Contributing

Rementum gives your agents a memory they can trust. Knowledge lives in linked, encrypted Markdown articles. Every canonical change is staged, versioned, attributed, and conflict-checked before it replaces the previous version, so two agents never overwrite each other's work.

Illustrated overview: agents read a compact index, stage writes, and share one versioned brain. See it live, drawn in your browser, on rementum.dev.

Why Rementum

• 🧠 Agent-first: agents read a compact routing index, then open only the article it points to.

• 📝 Staged writes: each proposal is checked against live canon before it lands; conflicts wait for review instead of overwriting.

• 🔒 Envelope encryption: article bodies use per-brain keys and position-bound AAD; titles and metadata stay searchable.

• 🔎 Hybrid search: Rementum fuses routing metadata, PostgreSQL full-text, and local multilingual embeddings.

• 🤝 Coordinated agents: leased tasks and maintenance proposals flow back through the same staged protocol.

• 📦 Yours to keep: self-hosted, open source, and exportable as Markdown whenever you want.

Related MCP server: Cortex

How it works

An agent never loads the whole brain. It reads a compact index, opens the one article it needs, and proposes changes through a staged protocol that checks for conflicts before anything replaces the current version.

flowchart LR
    IDX[Agent reads the routing index] --> ART[Opens the one article it needs]
    ART --> STG[Stages a write against a base version]
    STG --> CHK{Base still the current version?}
    CHK -->|yes| PRM[Promote]
    CHK -->|no| PRK[Parked as a conflict]
    PRK -->|reviewer resolves| PRM
    PRM --> VER[New immutable version + audit event]

Rementum stores titles, summaries, and bodies exactly as agents write them. Nothing is generated: the routing index is the title and one-sentence summary each write carries, and no article text is ever sent to an external model.

Quick start

You can evaluate Rementum locally on your machine without a domain, or deploy it to a Linux server for production with automated TLS.

Option 1: Local evaluation (no domain required)

Run the full stack locally in 2 minutes using Docker Compose:

git clone https://github.com/rementum/rementum.git
cd rementum
cp .env.example .env
docker compose up -d
./scripts/create-owner.sh owner@example.com "Owner"

Open http://localhost in your browser.

• Web dashboard: Sign in with the owner email and password you just created.

• Connect an agent: Open Teams, copy your workspace MCP URL (http://localhost/mcp/workspace/WORKSPACE_ID), and connect Claude Code, Codex, Cursor, or any MCP client.

Option 2: Production deployment (server with domain & HTTPS)

Deploy on a Linux server with automated HTTPS via Caddy and encrypted backups:

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

The interactive installer prompts for your domain (e.g. memory.example.com), generates cryptographic secrets, starts the stack, runs migrations, creates the first owner, and provisions TLS certificates. Update later with ./scripts/update.sh.

See the installation guide and operations guide for requirements, backups, and recovery.

Security

Rementum uses application-layer envelope encryption with searchable metadata. Article and version bodies are encrypted with AES-256-GCM using per-brain data keys sealed with position-bound AAD, wrapped by an instance master key that never touches the database or backups. Article titles, routing summaries, slugs, backlinks, and vector embeddings remain unencrypted in PostgreSQL so hybrid search works without client-side decryption; treat them as sensitive derived data.

Rementum makes no external AI requests and generates nothing: titles, summaries, and bodies are stored exactly as submitted.

Read SECURITY.md and the security checklist before you store private knowledge. Report vulnerabilities through the process in SECURITY.md, not a public issue.

Architecture decisions and FAQ

Looking for details on our architectural trade-offs, why we chose PostgreSQL over Git, token efficiency, AGPL-3.0 licensing, or privacy? Read the Architecture decisions and FAQ guide.

Documentation and contributing

The full docs are at rementum.dev/docs: configuration, backups, upgrades, security, and agent connections. For local setup and checks, read the development guide and CONTRIBUTING.md.

Status: active development toward a production beta. REST and MCP contracts are versioned. We don't promise backward compatibility until 1.0.

License

Rementum is licensed under AGPL-3.0-only.

Rementum is an independent network service. Running Rementum internally for your team or organization does not require open-sourcing your proprietary code or agent workflows. AGPL ensures that enhancements to the core platform itself remain open to the community.