SYSTEM 03 / HERMES INFRASTRUCTURE

Hermes Infrastructure Full-Stack Flow

Messages enter through platform adapters, become agent sessions, collect memory and skills, call a model, execute tools/MCP servers, persist evidence, and deliver results back to the origin.

Stack layers full path from input to output
01WhatsApp/Discord/API
02gateway adapters
03session DB
04memory + skills
05AIAgent loop
06model provider
07tool registry
08MCP servers
09terminal/files/browser
10cron delivery
Architecture graph

Control plane + data plane + trust boundaries

Hermes repo AGENTS.md, live config shape, and process scan covering gateway, WhatsApp bridge, personal_core MCP, GitHub MCP, Playwright MCP, cron/background services.

user Platforms

WhatsApp, Discord, API, CLI

edge Gateway adapters

normalize message + route session

data Context loader

AGENTS, memory, skills, profile

runtime AIAgent loop

model call, tool call, iteration budget

ai Model provider

chat/completions or provider adapter

guard Tool registry

schemas, toolsets, dispatch

runtime MCP servers

personal core, GitHub, Playwright, WhatsApp

output Local execution

terminal, files, browser, git

data State layer

SQLite sessions, logs, memory, cron

output Delivery

origin chat, scheduled output, media

Runtime sequencehow data moves
  1. 1. Ingress

    A platform adapter receives a message or cron tick and resolves the target session/channel.

  2. 2. Context

    Hermes assembles project instructions, memory, skills, recent session state, and tool schemas.

  3. 3. Reasoning

    AIAgent calls the configured model and loops through tool calls until enough evidence exists.

  4. 4. Execution

    Tool registry dispatches to local tools or MCP servers; results return as structured evidence.

  5. 5. Persistence + delivery

    Session DB, logs, artifacts, memory, and cron output store state; the gateway sends the final message back.

Data contractssource → transform → destination
User event

WhatsApp / Discord / API→gateway normalization→agent session

Task context

memory + skills + project files→prompt assembly→model call

Tool evidence

local tools + MCP servers→structured result messages→agent loop / session DB

Final output

agent response→gateway delivery formatting→origin platform

boundary

Platform boundary

Adapters translate platform-specific events into a single internal session format.

boundary

Reasoning/execution boundary

The model decides; deterministic tools perform and verify actions.

boundary

State boundary

Temporary task progress stays in sessions; durable preferences go to memory; artifacts stay on disk or target platforms.