AI@Work meetup — March 2026

Forge

A personal intelligence stack built on a Mac mini

BOTTOM LINE UP FRONT

I built a local infrastructure stack that gives my AI agent persistent memory, relationship awareness, and self-monitoring — all running on a $600 Mac mini with zero cloud compute costs.

▶

The problem

AI agents are stateless. Every conversation starts from scratch. They don't remember what you built yesterday, who you work with, or what you're trying to achieve. They can't tell the difference between something you said and something you were testing.

Most people solve this by pasting context at the start of every chat. That doesn't scale. I wanted an agent that wakes up already knowing my world.

▶

What Forge is

Seven services running locally on a Mac mini, all auto-starting, all self-monitoring:

Key-value store

Foundation layer

Document database

Schemas + indexes

Vector database

Semantic search

Graph memory

Relationships

Memory validator

Write gatekeeper

Cache

Fast lookups

Message queue

Async routing

Total cost: one Mac mini, one Claude Max subscription. No AWS, no GPU rentals, no API metering.

▶

How memory works — three layers

Vector DB — the scout

Semantic search across everything. You say "what was I stressed about last week" and it finds relevant context — not by keyword matching, but by meaning.

Graph memory — the detective wall

Structured relationships between people, projects, and tools. The agent doesn't just find information — it follows connections. "Alex manages the project that depends on the tool you built" is a graph traversal, not a search result.

Memory validator — the gatekeeper

Every write gets validated before it lands. No duplicate entries, no wrong dates, no edges pointing to things that don't exist. The agent can't pollute its own memory with bad data.

▶

The "aha" moment

I asked the agent to map everything connected to a career goal I'm working toward. It traversed the graph and came back with something I hadn't pieced together myself:

"There are two separate stakeholder paths converging on this role — one person owns the data quality side, another owns the operations side. The tool you built last month is the bridge that satisfies both of their priorities."

That's not a search result. That's structural reasoning across relationships the agent learned over weeks. No amount of context-pasting into ChatGPT gives you that.

▶

Self-monitoring

The stack watches itself. Every layer has a check:

Crons verify their own writes landed after every run

Heartbeat checks all 7 services every 30 minutes

Dead man's switch on an external VPS alerts me if the Mac mini goes dark

Weekly review surfaces unconfirmed graph edges for human verification

If my Mac mini crashes at 3am, I get a Telegram alert within the hour. The agent diagnosed its own monitoring gaps and proposed the fixes.

▶

What I'd tell my past self

Build what you need, not what's impressive. Two of my seven services sat idle for weeks until the graph layer gave them a purpose.

The agent will get things wrong. It confused test data with my real opinions. That's why the validator exists — trust but verify, programmatically.

Timezone bugs will haunt you. My nightly memory pipeline failed for four nights because UTC vs local time wasn't explicit. Always specify timezone.

Let the agent harden itself. The best monitoring improvements came from asking the agent "what are your blind spots?" — it identified gaps I hadn't considered.

▶

The stack at a glance

Agent	OpenClaw ("Entropy") — orchestrator, judgment layer
Builder	Claude Code — writes the infrastructure
Infrastructure	Forge — 7 local services on Mac mini
External	n8n on VPS — webhooks, monitoring, relay
Interfaces	Discord + Telegram
Total cost	Mac mini + Claude Max + $5/mo VPS

John Henry — github.com/RadFlavor

No CS degree. Self-taught. Built in 3 weeks.