Hermes AI

Overview

Hermes is a long-running personal project to build AI agents that hold up in everyday use: structured memory, tool integrations, MCP connectors and explicit approval flows for sensitive actions.

The goal is to shorten the distance between a decision and the action it triggers, without giving up human oversight on consequential steps.

Problem

Useful work in personal infrastructure is fragmented: half-finished notes, manual file moves, ad-hoc API calls. Routine tasks take more attention than they should.

Solution

Long-running agent sessions that can read context, call tools, and pause for approval before doing anything consequential. The agent does not own the decision; it executes a defined step inside one.

Implementation

Python core, MCP connectors for tools and services, structured per-session memory, JSON-defined approval flows for sensitive actions.

Run on the HZE Infrastructure stack; tool calls and outcomes are logged for review.

Results

Working agent sessions in daily use for selected tasks; clear separation between autonomous and approval-gated steps.

Technical Challenges

Keeping memory relevant across many sessions without leaking prior context into unrelated work.

Choosing which actions require human approval — easy to over-approve and lose the benefit, easy to under-approve and lose the safety.

Highlights

• Long-lived agent sessions with structured memory and recovery.
• MCP connectors for tools and external services.
• Human-in-the-loop approvals for sensitive actions.