Agentic AI

hermes-local-layer

My second brain and assistant, built around Hermes Agent.

Overview

This local operating layer was built around Hermes Agent with the goal of serving as my second brain and assistant. There are four main functionalities: knowledge vault, mental space offload, thought organization, and information discovery. These functions are delivered through six channels: wiki, journal, tasks, feed, repo-scout, and stock-screener, each with its own specific harness.

The harnesses are built to allow only very selective access across systems. For example, the feed can read signals from the wiki, journal, and tasks to recommend reading, but it doesn’t have write permission. I will briefly describe what each subsystem does below.

The GitHub repository and its publication inventory provide more detail about the project.

Upstream Hermes Agent CLI/TUI · messaging gateway · tools · cron · reusable skills
Local operating layer orient → validate → mutate → verify
Journal
capture, organize, and archive thoughts
Wiki
source-only ingest, knowledge graph, and search
Tasks
projects, priorities, recurrence, and reminders
Feed
recommendations using journal/wiki/task signals
Scout
read-only GitHub project discovery
Stocks
market-data screens, charts, and review pages
Homepage
navigation plus generated presentation pages
Messaging workflows commands · feedback
scheduled reports
Served sites home · feed
wiki · stocks
Task integrations two-way edits
calendar · reports

Safety: canonical registries · subsystem validation · atomic writes
documentation checks · leak scans · filtered publication

Figure 01 Local architecture: Hermes Agent, the operating layer, and its connected workflows.

Wiki

The idea of a self-improving wiki was partially inspired by Andrej Karpathy’s gist. An issue I often face is reading something, then stumbling upon a related topic months or years later and trying to recall what I know about it.

The wiki addresses this: I dump whatever I read and find useful into the knowledge base, while the agent categorizes and organizes it into a “wiki.” When I ask questions, the answers are based strictly on sources within my “brain.”

Figure 02 The interactive graph connects related concepts across the wiki.

Journal

Journal is essentially a more eloquent version of myself. It turns my unstructured, random thoughts into coherent sentences and organizes them by topic. When I need to revisit my ideas, I consult this channel, which recalls both the ideas themselves and the thought process behind how they developed.

Tasks

Tasks is my personal secretary. I tell the agent what needs to be done, and it organizes things into task objects. These objects are sorted into a database and linked to projects. The channel also connects to Notion and Google Calendar for visualization. At the end of each week, it recaps what I have accomplished.

Feed / Repo-scout / Stock-screener

These are the information discovery channels. The feed produces two daily digests with five articles each: three aligned with my interests (inferred from the wiki ingestion log, journal, and task log), and two exploratory articles (as far from my interests as possible) to expand my perspective.

Repo-scout finds GitHub repos that I may find interesting. Stock-screener uses a technical approach to highlight interesting charts, prompting me to look into companies I may not have otherwise discovered.