All insights
AI Product Building AI Agents Architecture

Sessions are runtime infrastructure, not just resumable transcripts

Hermes stores sessions in SQLite with search and lineage so CLI, messaging platforms, and scheduled jobs all attach to one session plane — routing can resolve before the model even runs

@aparnadhinak (Aparna Dhinakaran) — Hermes Harness Architecture · · 7 connections

Editor-first harnesses treat a session as a transcript you reload to resume. Hermes treats the session store as live infrastructure: sessions carry source tags, parent-child lineage, and routing metadata, so a chat message, a CLI invocation, or a scheduled job can all attach to the same session plane — and a message can be routed to the right session before inference even begins. It even exposes a model-facing search tool so the agent can recall across prior sessions mid-task rather than relying only on static injection. This is Treat an agent as an operating system, not a stateless function taken to its conclusion, and a concrete reason Memory is a harness responsibility, not a pluggable component; persisting the session plane is also the structural fix for how Reasoning evaporation permanently destroys agent decision chains when the context window closes, extending the everyday logic of Session capture turns ephemeral AI conversations into a compounding knowledge base into the runtime itself.