I Was Drowning in Opportunities. So I Built a Compass.

My Tech Stack

Hermes Agent (running locally via WSL/Ubuntu)

Gemini (connected to Hermes as the underlying model)

React + TypeScript

Vite

Tailwind CSS

Node.js

How I Used Hermes Agent

Hermes Agent is the reasoning engine at the center of Lodestone. Every meaningful feature runs through it.

Memory

Hermes stores and retrieves the user's North Star, motivations, and planning horizon across sessions. This means decisions logged today are evaluated against goals set months ago -- not just the last message in a chat window.

Multi-Step Reasoning

When generating reflections, Hermes doesn't evaluate a single interaction. It reasons across the full decision history, identifies trends, detects drift, and compares recent behavior to the user's intended direction. That multi-step process is what separates Lodestone from a chatbot with a scoring widget.

Planning

Hermes produces actionable next steps grounded in the user's actual history and long-term objectives. It doesn't just describe what happened. It generates a plan for what should happen next.

Agentic Workflow

The reflection flow makes the agentic behavior visible and legible to the user, with three explicit phases: memory, reasoning, and plan. This was a deliberate design choice. The goal was for users to see Hermes working, not just producing output.

Challenges I Ran Into

The hardest problem was the cold start, a compass is useless without a direction. Most productivity tools fail because they demand a perfectly articulated goal upfront. The insight that unlocked Lodestone was letting Hermes extract a North Star from natural, fragmented input. That single design decision made the whole system feel accessible instead of intimidating.

The second challenge was making the agentic behavior legible. It's not enough for Hermes to reason well internally; users need to see it happening. Designing the three-phase reflection flow around memory, reasoning, and planning solved that.

What I Learned

The most valuable part of Hermes wasn't text generation. It was the combination of memory, reasoning, and planning working together across time.

Building Lodestone changed how I think about what AI products can be. The future isn't assistants that answer questions. It's systems that help people maintain direction -- that hold the thread between who you said you wanted to be and what you're actually doing today.

What's Next

• Weekly autonomous reflections triggered without user input


• Proactive drift alerts before the drop are significant


• Voice-based goal discovery


• Calendar and task integrations


• Long-term goal forecasting across 6 and 12-month horizons

Most AI tools help people do more.

Lodestone helps people remember why they're doing it in the first place.

For students drowning in opportunities, that's not a nice-to-have. That's the whole game.