About
PurpleDirective started with a simple question: what would happen if you treated AI agents not as disposable chat sessions, but as an organization that learns and improves over time?
The answer turned into a multi-agent system where a COO agent triages every task, an advisory council deliberates on complex questions from structurally different perspectives, operational agents execute domain-specific work within defined guardrails, and a closed-loop learning system transfers knowledge from cloud inference to local models running on sovereign hardware.
The Stack
Cloud runtime: Claude via Claude Code. This is where complex analysis, multi-agent deliberation, and knowledge generation happen. Full agent identities with rich context.
Local runtime: Ollama via Purple-Directive: CLI. Condensed agent identities under 2K tokens, MCP tool integration, session persistence. Runs on a homelab server with an RTX 5090 for inference and fine-tuning.
Memory: SQLite with FTS5 keyword search and sqlite-vec semantic embeddings. Hybrid retrieval using Reciprocal Rank Fusion. Knowledge persists across sessions and transfers between runtimes.
Infrastructure: Raspberry Pi network appliance (DNS sinkhole, VPN gateway, access point), homelab server (inference, training, services), Cloudflare for DNS and tunnels.
Principles
- Sovereignty first. Your AI should run on your hardware, keep your data local, and work offline. Cloud is a tool, not a dependency.
- Compound learning. Every session should make the system smarter. Knowledge fragments, corrections, and patterns accumulate. The same mistake doesn't happen twice.
- Structured disagreement. Three advisory agents with different cognitive styles produce better analysis than one agent trying to be balanced. Deliberation has a hard stop to prevent drift.
- Audit everything. Every agent output is reviewed before it reaches the human. Every decision is logged with rationale. Trust is built through transparency.
- Details matter. The difference between a system that works and a system that works well is in the edge cases, the error handling, and the things nobody checks.
Open Source
The framework, protocols, memory kernel, and CLI tool are all public on GitHub. The goal is to demonstrate that a single person can build and operate a sophisticated AI agent system without enterprise infrastructure or a team of engineers.
If you're building something similar, the Violet repository has the full architecture — agent definitions, protocols, memory schemas, and operational templates.