The problem it solves
Every agent framework I tried had the same three failures. Agents forgot everything between sessions, so every day started from zero. They couldn't be trusted to run alone, so nothing happened unless I was watching. And one agent doing everything — plan, build, verify — collapsed under its own context the moment a task got real.
Hermes is the answer built backwards from those failures. It isn't a chatbot with tools bolted on. It's an operating layer: persistent memory underneath, orchestration in the middle, interfaces on top, and enough verification infrastructure that I can hand it hours of work and sleep through it.
Architecture
Cognitive memory, not a vector dump
The memory layer is a custom implementation of ACT-R cognitive architecture. Knowledge lives in chunks with activation levels — recency and frequency raise them, disuse decays them logarithmically, associative links spread activation through related concepts. Retrieval is a competition, not a similarity search: the most activated chunk wins, so well-earned knowledge stays sharp and noise sinks without anyone curating it.
It also produced my favorite class of bug: early on, a keying flaw collapsed thousands of unrelated memories into one black-hole record that everything silently merged into. Finding it required auditing retrieval against ground truth instead of trusting the health checks — which became a standing rule: memory systems must be audited against reality, not against their own telemetry.
Scheduled autonomy
A cron layer gives the platform its own sense of time. Jobs run without a user present: data collection, watchdogs, overnight builds. The hard-won lesson from running these: prompts for unattended work must be compiled upfront into small, verifiable steps — each naming its files and done-test — with fresh context per step, revert-on-failure discipline, and a budget cap. A 400-step instruction guide decays by hour six. Ten behavioral laws don't.
Multi-agent design rooms
For design problems, multiple agent instances deliberate in a structured conference room with locked contracts: they poll, respond continuously, and argue until the design converges. It sounds theatrical until you watch it catch a bad assumption in minute four that would have cost a weekend in hour twenty.
Verification as infrastructure
Nothing ships inside the platform on vibes. Work passes through staged gates: integrity checks first, then reproducibility, then adversarial probes aimed exactly where the step is risky. Failed builds revert automatically and get blocked, not retried hopefully. The gate stack has killed more of my own clever ideas than any reviewer.
What it enables
Fish Bite was specified, built, and regression-tested through this platform. The fishing forecast engine below runs on a schedule the agent itself maintains. This website was audited by it. That's the actual test of an agent platform: not a demo, but a body of shipped work it carried.
// field note
The jump from v6 to v7 wasn't more intelligence — it was better infrastructure around the same intelligence. Gates, memory hygiene, and scheduling beat prompt wizardry every time. Here's the full history of why.
Stack
- Interfaces: chat, voice (speech in, spoken responses out), scheduled headless runs
- Memory: custom ACT-R implementation — activation dynamics, spreading activation, decay
- Autonomy: cron-driven jobs, unattended build loops, watchdog reporting
- Verification: staged gates, champion replay, adversarial probes on risky steps
- Inference: hybrid — frontier API for reasoning, local GPU for execution