Joshua Stoner

What happens when you give an AI persistent memory and use it every day for a month. The discoveries, the failures, and the moments it started surprising me.

The shift from stateless conversations to persistent context—and what it actually feels like week by week. Stopped repeating myself. Then it started anticipating. Then it started noticing things.

Three memory layers, plain-file architecture, naming conventions, and the mechanisms that keep a growing system from turning into a junk drawer.

23 specialized agents, model-tier assignments, orchestration patterns, and the safety layers that keep autonomous execution from going sideways.

What a month of building revealed about AI as infrastructure, the gap between specification and reality, and why memory compounds while statelessness doesn’t.

Rules that fire on their own. A twin that drafts in my voice. And the bright line I drew around unreviewed external action.

The feedback loop that reads the AI’s own failure record and proposes structural fixes to the agents that underperformed—and why the judgment stays manual.

The hidden cost of context loss—status theater, the "where is that?" tax, tribal knowledge walking out the door—and what it looks like when you fix it.

The architecture: baseline creator pattern, context connector, write queue, three-layer density model. How the pieces fit together and why it’s built from plain files.

Governance, legal discovery risk, AI trust calibration, information overload, metric gaming, and the other unsexy things that determine whether this actually works.

What transfers when you scale a personal AI system to a team—trust model, concurrency, identity, information density—and what breaks. The stonerOS → TeamOS journey.

The real architecture of stonerOS—three memory layers, 23 agents, 21 hooks, plain markdown files. No database. No vector store. Two months of iteration documented honestly.

What happens when your personal AI system is better than the one your company gives you. BYOD happened because personal devices were better—the same forcing function is building with AI.

I pointed my AI memory system at my own phone data—browsing history, app usage, photos, reminders. The patterns it surfaced were things I already knew but had never quantified.

The gap between the people who build AI tools and the people who teach others to use them—and why the most important AI skill isn’t prompting, it’s evaluating.

A 7-track alt-R&B EP written and produced in four days via prompt engineering. All lyrics mine. The same skills that power stonerOS, applied to music. Listen on the page.

Can't sing. Don't play anything. Made a 7-track EP anyway. On the gap between skills and mediums—and how stonerOS closed the same gap for thinking.

Every L&D team built the same prompt library. Here’s what actually needs to change about how we teach people to work with AI—and why the cheatsheet was never the ceiling.

Five months between jobs. A memory system, two native apps, an EP, and a writing practice. What the output actually looks like when you stop optimizing for a manager and start building on your own terms.

The value isn’t in what the AI knows. It’s in what you discover when you have to explain yourself precisely to something that listens.

AI commoditized creation. Knowing what to build, and how it should feel, still costs something. The curation is the work the tool doesn’t do.

I shipped an iOS app using AI agents as the team—not “AI-assisted.” Twenty agents, zero merge conflicts, and bugs only a human on-device could catch.

A/B tested two feedback formats with 169 participants to diagnose a ceiling effect. Designed a Frequency × Impact matrix and a quadrant-based intervention framework for manager conversations.

Replaced a 10+ step manual Qualtrics process with end-to-end Google Workspace automation. One form submission triggers folder creation, template population, a 10-step Zapier sequence, and results delivery.

Three-part facilitation guide for managers to lead their teams through AI adoption—before a formal org ask existed. Built for psychological safety with three adoption pathways.

Manager development workshop closing a diagnosed gap in feedback quality. Built around the SBID framework with a paired workbook. Reached ~60–70 managers during significant org change.

Production AI memory system: 3-layer hierarchy, multi-agent orchestration across three model tiers, automated CI/CD via Claude Code hooks, and staleness surfacing via context injection. Self-initiated, continuously evolving.

Native macOS menu bar app in Swift/SwiftUI that tracks Claude Code usage in real time. Plan limits via OAuth2, weekly activity charts, API-equivalent cost analysis (53x ROI), and a 24-hour usage heatmap. Zero-config, light/dark mode.

klikWheel — iOS Music Player

Retro scroll-wheel music player for iPhone. Browse your local library or stream Apple Music through a tactile, haptic-click interface. Zero ads, zero tracking. Built end-to-end with multi-agent Claude Code orchestration.