meraGPT

June 17, 2026 · 6 min

Teach your agent to remember you

The thing a public chatbot can’t be is yours. You can talk to it all year and it still meets you as a stranger every morning. And the parts it does keep live on someone else’s servers, under someone else’s account, feeding someone else’s model.

meraGPT means “my GPT”. The whole point is an AI that is actually yours, and the thing that makes it yours more than anything else is memory: it remembers you, the notes live on your box, and nobody else can read them. This is how to set that up properly, in about ten minutes, with no SSH and no command line.

What memory actually is on your box

Hermes keeps its memory as plain markdown files under ~/.hermes/memories on your own machine. Not a vector database you can’t read, not a row in our control plane. Markdown. You can organise it into folders, the agent reads and writes it as you talk, and it carries across every session and every channel: the same memory whether you reach your agent from the browser, Telegram, or the desktop app.

Because the notes are files on your box, they are private to you by construction. We run the machine, but our control plane has no way to fetch your memories, and when you leave, /backup zips the whole thing so you take it with you.

Step one: tell it who you are

The single highest-leverage thing you can do is seed a profile. Open the CHAT tab on your dashboard and paste something like this (edit it to be about you, obviously):

Save this to memory under profile/about-me.md and
keep it up to date as you learn more about me:

# About me
- Name: Asha. Call me Asha.
- Work: solo founder, building a logistics SaaS.
- Stack I care about: Postgres, Next.js, Hetzner.
- Timezone: Asia/Kolkata. I start work around 9am.
- How I like answers: short first, detail if I ask.
  No hedging, no "it depends" without a pick.
- People: my cofounder is Ravi (backend). My
  accountant is Meena, year-end is in March.
- Ongoing: migrating off DynamoDB; learning German.

When something I say later contradicts or extends
this, update the file rather than starting a new one.

That is it. The agent writes the file, and from now on every conversation starts with it already knowing your name, your work, how you like to be answered, and who the people in your life are. No restart, no setup screen.

Step two: make it keep itself current

A profile you have to maintain by hand goes stale. Add a small skill so the agent updates its own memory of you as it goes. Skills are markdown files the agent reads before it acts; create one from the SKILLS tab (or just paste it into CHAT and ask it to save the skill):

---
name: remember-me
description: |
  Keep an accurate, evolving memory of the user.
  Use on every conversation.
triggers:
  - any message where I share a durable fact about
    myself, my work, my people, or my preferences
---

# Remembering the user

Maintain memory under profile/ as the source of truth
about who I am.

1. At the start of a session, read profile/about-me.md
   so you already have context. Don't announce that you
   did; just use it.
2. When I say something durable (a preference, a person,
   a project, a deadline, a decision), quietly update the
   relevant profile file. Durable means it'll still be
   true next week, not "I'm tired today".
3. Keep it tight. Edit existing lines instead of piling
   up duplicates. If two facts conflict, keep the newer
   one and note the change.
4. Never put secrets (passwords, full card numbers, API
   keys) in memory. If I paste one, use it and don't
   save it.
5. If I ask "what do you know about me?", read profile/
   back to me so I can correct it.

Save it. The agent picks it up on the next message. From here it gets to know you the way a good assistant would: you mention once that you prefer Hetzner over AWS, and three weeks later it just assumes it when you ask for a deploy plan.

The test that makes it click

Tell it something offhand in the browser (“Ravi’s out next week, so I’m covering backend”). Close the tab. Later, message your agent from Telegram and ask “who’s on backend right now?” It answers from memory, on a different surface, with no prompt from you. That moment, when it knows something you only said once, somewhere else, is the moment it stops being a chatbot and starts being your agent.

Why this is different from chatbot “memory”

It’s actually private. Public tools that remember you do it on their servers, tied to your account, often in scope for training. Your meraGPT memory is markdown on a box only you can reach. We can’t read it, and we don’t want to.

It’s portable. The notes are yours. /backup exports memories, skills and sessions as a zip. Leave whenever you like and the memory leaves with you. Nothing is held hostage.

It compounds. Once your profile is in memory, every other workflow gets smarter for free. Your inbox brief knows what matters to you. Your calendar prep can pull up what you saved about the person you’re about to meet. Memory isn’t a feature in a silo; it’s the layer the rest sits on.

What it costs

Almost nothing. Reading your profile at the start of a session is a few hundred tokens; updating it is a sentence or two. On a cheap model like openrouter/google/gemini-2.5-flash-lite it rounds to zero against what you already spend chatting. The memory itself is free; it’s just files on the box you’re already paying $9.99 a month to run.

Related