A multi-agent AI system that never forgets

The brief was to design and build a personal AI operating system: multiple distinct AI agents, each with a specific role, all drawing on one shared memory rather than starting from zero in every conversation.

The architecture runs on self-hosted infrastructure, n8n, a vector database and Postgres, all in Docker on a cloud VPS, ingesting events from more than ten sources continuously. Four distinct agents were built as genuinely separate entities sharing one memory backend, with a hard rule that knowledge is never silently deleted, only removed through a deliberate manual admin action. The whole system is reachable through a simple chat interface rather than a dashboard.

A working multi-agent memory system built from open infrastructure rather than a single vendor's black box, with a six-phase build plan guiding the rollout in stages.