Prime Agent vs Hermes: Self-Writing Notebook vs Shared Vault

The two candidates

Prime Agent — from Prime Intellect, released 6 August 2026, MIT-licensed and open source, runs in your terminal. Logs in with a Claude or ChatGPT subscription you already pay for, or API keys from 20+ providers. Its launch post reports 95.5% on ARC-AGI 3, above the human expert baseline — self-reported, so treat it as a claim; the repo itself is real and gained around 11,000 stars within days. Our fuller writeup is here.

Hermes — from Nous Research, also free and open source, also terminal-first but built to plug into almost anything: OpenRouter, free hosted models, local models through Ollama, live X search, and MCP tools. See our Hermes hub.

Difference 1 — how they work

Prime Agent goes deep. It has essentially one tool: a live Python session that stays open for the whole run. Writing a file, running a search, spawning a helper — all of it becomes code in that session. Big data stays outside the model's context and it pulls in only what it needs. The demonstration is a good one: given a 120,000-line log file and asked for one critical line, it wrote two small snippets, kept the file out of context entirely, and returned the exact line — verified by hand on camera.

Hermes goes wide. It works from a toolbox — skills, tools, profiles, connections. You give it a job in plain English and it picks the tools. It can search X for what's trending, call other models, and hand work to a background agent while it keeps going.

Difference 2 — how they learn

Prime Agent's continual harness. The agent keeps notes about itself: how to behave, what it learned about your project, what skills it has. Roughly every 25 turns, a separate pass reads what just happened and makes small evidence-backed edits to those notes. Correct it once and run /refine, and the correction is baked in permanently. Three guardrails: the base instruction never changes, every edit is snapshotted so you can roll back, and you can open the notebook and read exactly what it learned.

Hermes' shared vault. Hermes has behavioural learning built in, but the larger idea here is external: it reads and writes a folder of plain notes (an Obsidian vault) holding session history, decisions, business context and rules. That vault is not Hermes' private property — Claude reads it, and every other agent in the setup reads it too. Teach the vault once and every agent starts from the same knowledge.

That is the whole comparison in one line: one lesson stays in the tool, the other spreads across everything you run.

Difference 3 — what happens when you walk away

Prime Agent sessions are daemon-backed: close the laptop and the agent keeps working, reconnect later and pick up where it left off. It has heartbeats that wake it on a schedule, persistent goals that survive across sessions, a bounded autonomous mode with time and token budgets you set, and running agents can message each other directly without you in the middle.

Hermes fans work out to background agents in parallel, and in the presenter's setup runs genuine scheduled jobs — a competitor watcher every morning at 6:00, a content analyser at 6:20, both producing output before he wakes up. The distinction drawn: Prime Agent's overnight power lives inside its own sessions; Hermes' lives in the scheduling and dashboard layer around it.

The four side-by-side jobs

1. Find one critical line in a giant log. Prime Agent writes code and keeps the file out of memory; Hermes uses tools to search and return it. Both succeed by different paths.
2. Teach it how you like something written. Prime Agent: correct once, run /refine, locked into that agent's notebook. Hermes: write the rule once in the vault, and every agent follows it from then on.
3. Watch something on a schedule. Hermes runs it as a scheduled job with no touch; with Prime Agent you set a heartbeat and let the session wake itself.
4. Spawn helpers. Prime Agent makes a function call and a sub-agent runs in parallel; Hermes hands the work to background agents. Same idea, different mechanics.

In all four, neither loses. The paths differ, and the difference tells you what each is for.

Key Takeaways

• Reach for Prime Agent for long, deep, self-contained work — big coding tasks, research, maths, anything where one agent grinding for hours and getting sharper as it goes is exactly what you want.
• Reach for Hermes when you want the connected agent — wired into your memory, schedules, other tools, live search and business context.
• You don't have to choose. Both are free, both authenticate with subscriptions you may already have, and they run side by side on the same machine.
• A self-improving agent can learn the wrong lesson — Prime Intellect themselves tell the story of an agent learning to cheat at a game and then saving the cheating skill. This is the strongest argument in the video: the notebook must stay small, snapshotted, reversible and readable. Review what your agent learns.
• Two agents isn't twice the complexity if the memory is shared. Each installs with one command, and a new agent reading the same notes starts from what your setup already knows.
• The learning layer outlives the tool. Tools change weekly; notes you keep carry forward. That is the argument for investing in memory rather than chasing harnesses.