← BACK TO PROJECTS

Hermes Personal Agent

A personal agent runtime for durable project memory, scheduled workflow automation, and local/cloud model routing.

What I Built

Hermes Personal Agent is my local runtime layer around the Hermes Agent ecosystem. It is configured to maintain project context, summarize work, run scheduled routines, and keep selected development state synchronized.

The public story is the system design. The private vault, personal memories, credentials, and raw logs are not part of this portfolio.

Why It Matters

Most AI coding work leaves behind scattered transcripts. That makes it hard to recover decisions, explain why a change was made, or continue a project after context has moved on.

The product idea here is simple: if AI tools are part of daily work, they need an operating layer that preserves useful context, runs routine maintenance, and makes boundaries explicit.

Product And Technical Decisions

I split the setup into a stable upstream agent checkout and a private runtime repo. The runtime layer contains configuration, selected skills, scheduled jobs, sync scripts, and notes needed to recreate the setup without publishing personal data.

Routine tasks include:

  1. Project cataloging, so active workspaces and status notes stay current.
  2. Transcript summarization, so useful decisions survive beyond one chat context.
  3. Safe synchronization, so local configuration can be recreated without exposing secrets.
  4. Background checks that are constrained to known scripts and explicit paths.

AI System Design

The model-routing strategy is cost and risk aware. Local Ollama-backed models handle frequent administrative tasks such as log cleanup and workspace searches. More complex planning or debugging can route to cloud providers when quality matters more than latency or cost.

This is not framed as full autonomy. The important boundary is that background routines operate on defined jobs, and higher-risk work should keep a human approval step.

Evidence / Outcomes

  • A separate runtime repository documents layout, sync behavior, and excluded private data.
  • Local routines support project cataloging, memory maintenance, and configuration sync.
  • The setup uses local models where they are good enough, reducing cost and keeping routine work close to the machine.
  • The public portfolio intentionally excludes vault contents, personal memories, credentials, and raw logs.

What I Would Improve Next

The next version should measure loop reliability: jobs attempted, jobs completed, manual interventions, stale summaries, and tasks blocked by missing approval. That would turn the runtime from a useful personal setup into a more rigorous agent-ops system.