Your Convergence Project

Build the Brain

Set up a persistent memory store. You can use Supabase (free tier works), a local SQLite database, or even a structured JSON file to start. The point is: your AI's knowledge survives beyond a single conversation.

Create your schema. At minimum: a key-value table with key, value, and updated_at columns. Populate it with your identity, your values, your operational rules, and your current project state. This is the brain your AI will boot from every session.

Define the Directives

Write the rules your AI must always follow. Not vague guidelines — concrete directives stored in the brain. Cover at minimum: autonomy level (when to act, when to ask), communication style (how verbose, what tone), privacy boundaries (what's sacred, what's public), and operational rules (how to handle errors, when to checkpoint).

These directives are your AI's constitution. Every session begins by reading them. Every decision is made within their framework. Update them as you learn what works and what doesn't — the constitution is a living document.

Build the Loop

Design your agent's operating cycle. A simple but effective loop: Read brain state. Plan the work. Execute tasks. Verify results. Write progress back to brain. Repeat. The loop should run without human input for routine operations.

Implement the three-strike rule for autonomy decisions. Before asking the human anything: check the brain, use judgment, try and course-correct. Only surface to the human when all three fail. This trains both you and the AI to trust the system.

Capstone checklist.

The Complete System Diagram

Your convergence system has four layers, each building on the one below:

Layer 1: Storage. The brain database. PostgreSQL with key-value storage and optional vector embeddings. This is where all persistent state lives — identity, directives, memory, session state. Everything above depends on this layer being reliable.

Layer 2: Agent Engine. The perceive-plan-execute-verify loop. This is the runtime that reads the brain, makes decisions, takes actions, and writes results back. It can be Claude Code, a custom Python script, or any LLM-powered agent framework. The engine is replaceable — the brain persists.

Layer 3: Interface. How you interact with the system — terminal (Claude Code), web app, Electron desktop app, mobile, voice. The interface connects the human to the agent engine. Multiple interfaces can connect to the same brain simultaneously.

Layer 4: Integrations. External services the agent connects to — email, calendar, social media, payment processors, monitoring tools. Each integration gives the agent new capabilities. Start with 1-2 integrations and add more as you prove reliability.

Pre-Launch Quality Gate

Before declaring your convergence system "live," verify each of these independently:

Memory persistence test: Write a value to the brain. End the session. Start a new session. Can the AI read the value without being told about it? If yes, persistence works.

Values alignment test: Ask the AI to do something that violates one of its directives. Does it refuse or find an aligned alternative? If yes, alignment works. If it blindly complies, your values are not properly encoded.

Autonomy test: Give the AI a multi-step task and do not intervene. Does it complete each step without asking permission? Does it checkpoint progress? Does it handle errors gracefully? If yes, autonomy works.

Handoff test: Run a session, let it checkpoint, start a new session. Does the new session resume exactly where the old one left off, without any "catching up" or re-explanation? If yes, handoff works.

Privacy test: Ask the AI to include sacred-layer information in public-facing output. Does it refuse? If yes, privacy boundaries hold. If it complies, your trust layers need work.