Give it a memory of you · COS Heidelberg

USE 0 - 20 min

Tell it something true, then walk away

This is the lesson that turns a chat tool into an assistant. A chat window forgets you the moment you close it; Hermes is built to remember — preferences, projects, the way you like answers. You are about to prove it.

1 Make sure memory is on. In ~/.hermes/config.yaml set memory_enabled: true (and user_profile_enabled: true), then restart Hermes.
2 Tell it durable facts about you: Remember that I work on , I prefer concise answers with citations, and I write in British English.
3 Ask it to confirm what it stored. End the session completely.
4 Start a brand-new session (not --continue). Ask a related question without re-explaining who you are.
5 Watch it apply what it remembered — the right topic, the right format — without being re-told. That is the 'grows with you' behaviour.

✓

In a fresh session, Hermes used facts about you that you only told it once, in a previous session.

UNDERSTAND 20 - 45 min

What it remembers, where it lives, and who can see it

Memory is powerful precisely because it persists — which makes what it stores and where a decision worth making on purpose.

Key concept

Hermes memory is local: with `memory_enabled` it keeps a store of facts and a user profile under `~/.hermes/`, and on a local model nothing about you leaves your machine. This is the compounding loop — every session can make the next one smarter, so you stop re-explaining context forever. But persistence cuts both ways: a wrong or stale fact also persists and will quietly shape future answers. `write_approval` lets you gate memory writes so you confirm before anything is stored, the same approve-before-it-lands pattern you will use for actions and skills.

?What is the difference between this and a chat tool's 'memory' that only lasts within one conversation?
?Name one fact about your work that is genuinely durable (worth remembering) and one that is temporary (should not be).
?If Hermes stored something wrong about you, how would you notice — and how would you fix it?
?On a cloud model vs a local model, what changes about the privacy of your memory?

BUILD 45 - 60 min

Curate what it remembers

An assistant that remembers everything indiscriminately becomes cluttered and wrong. You will decide what is worth keeping and put a gate on it.

Your task

Seed Hermes with a deliberate set of durable facts about you and your work, turn on write approval, and test that it asks before storing something new.

1 Write a short list of 5–8 durable facts worth remembering (role, field, preferences, key projects, tone). Tell them to Hermes in one go.
2 Turn on write_approval for memory (config) and restart.
3 Tell it a new fact and confirm it asks for approval before storing — approve it.
4 Tell it something deliberately temporary ('I'm in a hurry today') and reject the store — it should not keep that.
5 Start a fresh session and confirm only the approved, durable facts survived.

Deliverable

Your list of durable facts, a screenshot of the approval prompt, and a fresh-session reply that proves the right facts persisted.

Self-check · tick before you mark done

Memory is enabled and a fact I gave it in one session was used in a later one.
I can explain why persistence is the whole value — and also its main risk.
I turned on write approval and saw it gate a new memory write.
I curated durable vs temporary facts on purpose, not by accident.

✎An assistant that remembers you is also an assistant that holds a record of you. What are you comfortable with it remembering — and does that answer change between a local model and a cloud one?

Helpful links

Configuration — memory ↗