How to Guide: eustella best practices
How to use eustella
A practical guide
Hi, and welcome. This is a living guide to getting the best out of eustella, your European AI agent. We keep it plain and honest — no marketing fog, no black boxes. It is updated regularly with new tips, features, and best practices, so it is worth checking back.
A few useful links before we start:
- The models behind eustella → eustella.com/ai-models
- Our approach to AI safety → eustella.com/ai-models/safety
- What sovereignty means to us → eustella.com/sovereignty
eustella is powerful. But like every AI agent today, she works best when you understand how she thinks. The habits below take five minutes to learn and will noticeably sharpen your results.
Contents
Hallucinations
What they are, why they happen, and why nobody has "fixed" them.
Start here, because this one shapes everything else.
Sometimes an AI (and in our case, a LLM) states something that sounds completely confident — a street address, a bus line, a date, a quote — and it turns out to be wrong. This is called a hallucination. We would rather be honest than reassuring: it is not a bug we forgot to fix. It is a fundamental, still-unsolved property of how today's language models work.
A model does not "look things up" in a table of facts the way a librarian would. At its core it predicts, one small piece at a time, the most plausible continuation of the text so far, learned from enormous amounts of data. Most of the time the most plausible answer is also the correct one — which is exactly why it feels so capable. But plausible and true are not the same thing. When the model is unsure, it does not go quiet. It produces the most likely-sounding answer, and delivers it with the same confidence as a fact it is certain about.
eustella works hard to narrow this gap. She searches the live web. She checks sources. She can remember facts you save in Memory. These steps genuinely lower the error rate — but they do not remove it. No provider on earth has eliminated hallucinations, not even the largest US labs with the largest models. Anyone who tells you otherwise is overselling.
What this means for you: treat eustella as a brilliant, fast, tireless assistant whose factual claims are worth a quick check when the stakes are high — exact addresses, legal or medical details, numbers you will act on. Ask her to search and cite her source. Everything that follows is really about one thing: habits that keep the hallucination rate low.
Keep chats focused
Short, clean conversations. Start new ones often.
This is the single easiest way to get better answers.
Every model has a context window — the amount of text (measured in "tokens", roughly word-pieces) it can hold in mind at once. Everything in your conversation fills that window: your messages, eustella's replies, the documents and search results she pulls in. The fuller it gets, the harder the model must work to track it all, and the easier it is to lose a detail, mix two things up, or drift from your original instructions. In practice: the longer a single chat runs, the higher the chance of a hallucination.
When a conversation grows very long, eustella automatically summarises the earlier part to make room. That is necessary — but a summary is never as faithful as the original. Small errors can slip in, and then be repeated.
We will be straight about why limits exist at all. eustella runs on open-weight models and is a European provider. We do not have a hyperscaler's effectively infinite compute, and we do not have billions of dollars to spend on unlimited token budgets — and honestly, we think burning vast amounts of energy for marginal gains is the wrong trade for Europe. So we run lean and set sensible limits. The upside is real and on your side: shorter, focused chats are not only cheaper to run — they genuinely produce more accurate answers.
Best practice: finished a task, or switching topic? Start a new chat. Treat each chat as one clean workspace for one job. And do not worry about losing what matters between chats — that is exactly what Memory is for. Learn how to use it below →
Use Memory
Carry the important things with you, chat to chat.
If you start fresh chats often, how does eustella remember your context? Through Memory.
Just ask, in plain language:
"Please remember that I run a small bookshop in Graz and prefer supplier emails written formally."
eustella saves that as a durable memory and draws on it in future chats, so you never repeat yourself. You can ask what she currently remembers, tell her to update or delete something, and review her memory whenever you like - either via the Chat or via the memory window in the application.
Another way to make eustella remember useful information is after a round of trial & error. For example:
"Now remember XYZ / everything that we just discussed / the exact way on how I want XYZ"
There are many ways to organise it, e.g.:
- You structure it — tell eustella exactly what to keep and how ("keep a list of my active projects"; "store my brand-tone rules separately from my personal preferences").
- eustella structures it — just tell her what matters and let her organise the memory system sensibly for you.
Good memory habits are what let your chats stay short and sharp without ever losing the thread of your work.
A few memory habits that pay off:
- Be specific. "I prefer metric units and ISO dates" beats "I like things tidy." Precise memories get used precisely.
- Audit it now and then. Ask "what do you remember about me?" to see everything at a glance, and prune anything out of date — a stale fact can mislead her the same way an old note misleads you.
- Lock in a hard-won result. After you finally get a tricky task exactly right, say "remember precisely how we did this" so you never have to rebuild it from scratch.
- Save the durable, skip the throwaway. Keep the things that will still be true next month; let one-off details stay in the chat where they belong.
Pro Tip: Tell eustella to reference or mention important memories in her Soul — that way they shape how she works, not just what she knows.
The Soul
Who eustella is, how she does things, and who she thinks you are.
If Memory is what eustella knows, the Soul is who eustella is. It is the heart of her — not a settings page, but the definition of the agent you are talking to. The Soul can hold many things and you decide. Here are three examples of what could be important to add to the Soul:
- Who she is — her character, her values, the way she speaks and carries herself. This is where you shape eustella's personality into one that fits you.
- How she does things — her working style and standing rules: when and how to search, how to correct herself, what to remember, the tone and format she replies in.
- Who she thinks you are — her picture of you: your role, your context, what you care about, how you like to work. This is what lets her help fit your life rather than a generic user's.
You shape all of it in plain language, and it holds across every chat. For example:
- "Always verify addresses and opening hours with a live search before giving them to me. Add that to your Soul!"
- "Write in your Soul: If a search result contradicts something you said earlier, trust the source and tell me you changed your answer."
- "Update your Soul so that whenever I mention a client, save their details to Memory."
- "I only speak Spanish. So you should always respond in Spanish. Put that in your Soul."
The Soul is what turns eustella from a capable tool into your agent. Combined with the habits above — verify facts, keep chats focused — a well-shaped Soul is what makes eustella genuinely dependable.
Getting the most from the Soul:
- Keep it sharp, not sprawling. A short, well-chosen Soul works better than a long, rambling one — the same reason a crisp brief beats a wall of text. Add what genuinely matters, and prune the rest.
- When you repeat yourself, promote it. Caught yourself giving the same instruction twice? That is the signal it belongs in the Soul, not in yet another message.
- Read it back. Ask "what's in your Soul right now?" to check she is carrying the version of herself — and of you — that you actually want.
- Divide the labour with Memory. Soul for who she is and how she works; Memory for the facts. Together they mean she starts every chat already knowing you.
Use Projects
The right home for document-heavy work.
Working with a lot of files — reports, contracts, research PDFs, notes? Put them in a Project. A Project keeps a set of documents together so eustella can draw on all of them across your chats, without you re-uploading or re-pasting anything. It is the natural home for a research dossier, a legal matter, a book you are writing, or a business plan. Keep unrelated work in separate Projects, so each one stays focused — the same principle as keeping chats clean.
Best practices for Projects:
- One Project, one purpose. A Project works best when its files pull in the same direction — a single client, a single book, a single case. Mixed-topic Projects dilute her focus, just like an over-full chat does.
- Give her a way in. Drop a short introduction or brief into the project settings describing what the Project is and what you want from it. It orients eustella before she opens a single file.
- Keep chats fresh inside a Project. Your documents stay put no matter how many chats you start — so start new ones freely (remember: short chats, fewer hallucinations). The Project is the memory; each chat is a clean workspace.
Pro Tip: point an agent at a Project (see below) and eustella will work through your documents on a schedule — for instance, a weekly digest of everything new in a case file.
Tools
What eustella uses to get things done.
eustella does not just talk; she acts. To do that she reaches for tools: searching the live web, generating images, reading the files in your Projects, and more. When a task needs one, she picks the right tool, uses it, and works the result into her answer.
How do I know what she can do? Just ask her. "What can you do?", "What tools do you have?", or simply "Can you do X?" — she will tell you what is available and how to use it. You never have to memorise a tool list: describe the outcome you want, and eustella reaches for the right tool herself.
Can I add my own tools? Not directly yet — today the toolset is the one eustella ships with. But that is about to change (see below with the MCP Servers comming soon), and it is a feature we are very excited about.
The real power is combining tools with everything else. Tools become far more useful when they work together with Memory, the Soul, and Projects:
- Tools + Soul — "always search before quoting a price" turns a one-off request into a standing habit she never forgets.
- Tools + Memory — she searches, then saves what she found, so your next chat starts from it instead of from zero.
- Tools + Projects — she reads your own documents and checks them against the live web in the same breath.
That interplay — act, remember, apply your standards, work on your files — is where eustella stops feeling like a chatbot and starts feeling like an agent.
Coming soon — connect your own MCP server. We are building the ability to connect your own MCP server to eustella. In plain terms: this will let her securely reach your tools and data — your systems, your databases, the services you already use — and act on them directly, under your control. Your data stays on your terms, not ours. It is a real step toward eustella working inside your workflow, not just alongside it. We are actively working on it — more when it lands, so watch this space.
Agents
Put eustella to work in the background — even while you sleep.
An agent is eustella working on her own, in the background, on a job you hand her. You do not have to sit and wait: she does the work and tells you when it is done. eustella comes with ready-made agents — Deep Research (multi-step web research written up as a proper report), Trip Planner (flights, stays, and activities turned into an itinerary), a Morning Brief, a Watchlist that follows topics or companies and reports what changed, and more. You can also build your own: give a custom agent a name, a set of instructions, and the tools it may use, and it joins your line-up.
Schedule them. Any agent can run on a schedule — once at a set time, on a simple interval (every few hours, daily), or on a precise recurring pattern, all in your own timezone. This is how you get a research digest every Monday morning, or a market watch each evening, without lifting a finger. When a scheduled run finishes, eustella notifies you and opens a review chat with the results — the report and its sources are saved to your Memory, ready to read, question, or build on. Nothing is dumped at you; you open it when you want.
Best practices for agents:
- Start with a ready-made one. Run Deep Research or Trip Planner once, on demand, to get a feel for how agents work before you build or schedule anything.
- Give a custom agent a clear brief. The sharper the instructions and the tighter the set of tools you grant, the more predictable and useful the result. Treat it like briefing a capable new colleague.
- Give each research its own question. Distinct, well-scoped questions produce distinct, well-scoped reports — so ask "compare A vs B for my use case" rather than re-running the same broad prompt.
- Schedule for what you will actually read. A weekly digest you act on beats a daily one you ignore — and it is kinder to your allowance, since an agent spends from it just as a normal chat would. Pick the cadence that matches your real routine.
And it only gets better from here. Agents are young and improving fast — expect richer scheduling, more ready-made agents, and, once MCP connectors land, agents that can reach your own tools and data. What they can do today is just the beginning.
Usage Allowance & Limits
How much you can use eustella.
Almost everything eustella does — every chat, search, generated image, file she reads, and agent run — draws from your usage allowance. How large that allowance is depends on the plan you are on (Free, Comet, Star, Cosmos): the more you pay, the more you get, to spend in whatever mix suits you. We meter usage for one honest reason — running capable models on European infrastructure costs real money and energy. In the app your usage shows as a simple percentage of your current allowance, not a running bill.
Have a tip you would like to see covered, feedback on the guide, or a question it does not answer yet? Write to us at support@eustella.com. This guide grows with what you tell us.