In this guide to the Hermes memory system, here’s what we’re going to cover: the five layers, the read-work-write loop that makes memory compound, and the galaxy view that finally makes people maintain it.
Short answer
Five layers: session context, persistent memory, skills, persona, profiles.
The loop: read before work, write after — agents maintain it, so it compounds.
Session context dies with the window; the persistent layer survives everything.
Audit with /journey weekly; export makes the whole system portable.
The Hermes memory system, layer by layer
Layer
What it holds
Lifespan
Session context
The current conversation and tool results
Dies with the session
Persistent memory
Your business, decisions, history — in the provider
Permanent, and portable
Skills
Solutions the agent wrote down as markdown
Permanent, reusable
Persona (SOUL.md)
Who the agent is and how it behaves
Permanent, per profile
Profiles
Isolated memory spaces per agent/job
Permanent, separable
Most disappointment with agents comes from conflating the first two rows. Session context is big and fast and gone the moment you close the window. The persistent layer is what survives — and it’s the layer most people never set up.
How the Hermes memory system moves information
Read before work: the agent opens with your context instead of a blank slate.
Work: the session does its job in context.
Write after work: what happened gets distilled back into the vault as notes.
Learn: hard-won solutions become skill files the agent keeps.
Compound: the next session — any agent, any interface — starts one step further along.
That write-back loop is the difference between memory as a feature and memory as an asset. A vault only you maintain rots; a vault your agents maintain compounds.
🔥 Want this set up without the guesswork? Seeing the memory system visualised is what makes people finally maintain it. Inside the AI Profit Boardroom you get the Agent OS as a ready-to-install file, a 30-day roadmap, daily tutorials the same day new tools ship, and four live coaching calls a week where you share your screen and get unstuck. → Get access here
Seeing it: the memory galaxy
The vault can be rendered as a 3D galaxy — every note a star, every link a constellation, recently updated notes glowing brightest. You fly through it, click a star, and you’re in the note; your agents read the same galaxy your eyes do.
It sounds like decoration and isn’t: the clusters show your real focus areas, the glow shows your now, and messy clusters are usually where the good ideas hide. You don’t need to tidy it — you need to see it.
The properties that make the system work
Per-profile isolation: each profile’s memory is its own folder, so experiments never pollute your working agent.
Shared provider across agents: one vault readable by every harness you run — learned once, known everywhere.
Auditability:/journey shows every memory and skill on a timeline, editable and deletable.
Portability: export packages the whole memory system into one file, credentials stripped — the identity survives machines and models.
A read-work-write loop over a persistent provider: the agent reads your vault before working, writes distilled notes back after, and keeps hard-won solutions as skill files — so every session starts further along.
What’s the difference between context and memory?
Session context is the current conversation and dies with it. Persistent memory lives in the provider on disk and survives sessions, machines and model swaps.
Where are skills in the memory system?
Skills are memory of how — solutions written as markdown files the agent reads before similar work. Notes are memory of what.
Can two agents share the memory system?
Yes — point them at the same vault. Per-profile isolation still applies where you want separation.
What is the memory galaxy?
A 3D visualisation of the vault — notes as stars, links as constellations, recent activity glowing — so you can actually see what your agents know.
How do I audit what it has learned?
/journey gives a timeline of every skill and memory, and you edit or delete anything wrong right there. Weekly is the habit.
The bottom line on the Hermes memory system
The Hermes memory system is five layers and one loop: read before work, write after, keep solutions as skills, isolate by profile, share by provider. Set the persistent layer up once, audit it weekly, and the same download quietly becomes a smarter agent every month.
