Scenarios

One system, three scales. Store, verbs, version history, and MCP surface are identical at every tier. Only reach and identity change.

Personal: your memory

A server of your own, on your machine or a remote host you join from every device.

• Agent memory: persistent memory for Claude Code, OpenCode, pi, and other MCP agents
• Personal knowledge base: local wiki, full history, TUI browser

Team: a shared world

One server a group reads, a reading room for humans, MCP for agents. Capability tokens per person, or the appliance’s built-in login.

• Five-minute appliance: world, broker, library, login, and HTTPS in one command
• Team knowledge base: the manual build with per-person tokens
• Public hub: public server, Let’s Encrypt, open reads

Enterprise: a knowledge system

Many worlds behind one broker, your identity provider in front. Teams join with one command; the broker mints per-world tokens. The indexing agent keeps the cross-world graph current.

• Organizational knowledge system: architecture, reference deployment, how users join
• The library and the librarian: human front-end over the universe

They compose

Anyone can run a personal memory and join the org system in the same agent, in Claude Code, OpenCode, or pi. All three plugin pairs share one ~/.demarkus state. /promote lifts a ready note from the private tier to the shared one through a curation gate; /memory-refresh pulls the authoritative copy back as it changes. The two plugins partition by server scope.

For the full agent-driven flow, how data lands in memory and gets promoted up to the knowledge system, see How knowledge flows.