Discord AI Bot Memory: What Your Server's AI Actually Knows

What is per-server AI memory?

Most AI support bots work from one knowledge base shared by every customer, or from whatever the model already knew about the world. Per-server memory is the opposite arrangement. Your community gets a store of its own, it starts empty, and it fills up with exactly what your staff decide to put in it.

That distinction sounds small until a member asks a question only your server can answer. "How long is the whitelist queue?" has no general answer. "Do I lose my rank if I leave for a month?" has no general answer. A shared FAQ engine either guesses or gives up. A per-server memory answers, because someone on your team told it once.

That is also the trade. An empty memory answers nothing useful on day one. The first hour you spend teaching it is the hour that makes it worth having, and there is no way around that hour.

What does the AI actually see when it answers a ticket?

Three inputs, in this order, and nothing else.

If the ticket opened through a form, the questions and answers are attached to that first message too, so the AI reads them before it says anything. That is worth knowing when you design your form: a field you add is a field the AI gets to use.

The label is deliberately thin. The category name reaches the AI and the category's description does not, which is a good argument for naming a category the way you would explain it out loud. There is more on that in our guide to ticket categories and to form questions.

What the AI cannot see (and what it has to ask for)

This is the half that gets assumed wrong, so it is worth being blunt. None of the following reaches the AI:

Adding the bot to a server absorbs nothing. If the AI knows your rules, it is because somebody taught it your rules.

There are three things it can go and fetch during a ticket, and each one has a condition on it:

All three are evaluated from the opener's point of view. If a member cannot see a channel, the AI cannot read it on their behalf, which means it can never quote a staff-only channel back to somebody who was not meant to see it. That is the same permission model the rest of the bot uses, covered in Discord ticket bot permissions.

How does a fact get into memory?

Three ways, and all three are somebody choosing to do it.

The dashboard mirrors the first two: a Train mode in the AI chat, and an Improve button on a closed ticket. Same permission gate on every surface, which is Manage Server or a staff role you have given training rights to. A ticket used from Discord shows as already used in the dashboard, and the other way round.

What you type is not stored word for word. It is rewritten once into a structured entry, the structured entry is what gets saved, and your original phrasing is discarded. This matters for two reasons: the AI reads a clean entry rather than your typo, and there is less raw text sitting in storage than you might expect.

The full instructions live in how to train an AI ticket bot, and the closed ticket route has its own walkthrough in the AI that learns from your real tickets. Two things about that second route are worth repeating here, because they are the ones people assume wrongly:

A ticket the AI resolved on its own is refused, because there is no better answer inside it. A ticket with no staff reply is refused too, because the only substantive text would be the AI's own mistake.

What happens when two people teach it different things?

Contradictions resolve rather than accumulate. When something new arrives that says the opposite of something already stored, the newer statement replaces the older one instead of both sitting there for the AI to pick between. The older version is not deleted, it is marked as replaced and stops being used, which is why correcting a policy is a one line job rather than a hunt for the wrong version.

Near duplicates are caught before they are written at all. If a staff member teaches something that is already in memory in slightly different words, the write is blocked and the existing entry is updated or superseded instead. You do not end up with four slightly different versions of the refund policy because four people taught it.

Instructions and facts are also kept apart, which is subtler and more useful than it sounds:

Teaching a rule as if it were a fact is the most common way a training session quietly fails, because a fact only surfaces when the question happens to match it, while a rule is in front of the AI every time it speaks. The system routes them for you and warns when something looks mis-filed, but knowing the difference makes your teaching sharper.

Can it read our website and our documents?

Yes, and more than most people expect. Two intake routes:

In training mode every text document you attach is ingested rather than skimmed. The AI is not given the option to read a file and reply "got it" without saving anything, which was a real failure mode before that was made deterministic. Long documents keep their section headings, so a question later reads the relevant section rather than the whole handbook.

Size never fails a job. A large file simply costs more of your token budget, and the job stops cleanly when the budget runs out. Nobody should ever be told to shrink a file.

There are genuine gaps here and they are worth stating plainly:

Does it look things up on the web?

It can, and the ordering is the important part. Anything about your products, your prices, your policies or your procedures must come from your own memory. Web lookup exists only for current outside-world information that memory cannot contain, and it is capped at three searches inside a single reply.

It runs on Brave Search, which is disclosed as a sub-processor in our privacy policy, and only the search terms leave, never the ticket conversation. It is a per-server switch, on by default, and it is metered against the same token budget as everything else rather than billed separately.

How much can one server's memory hold?

The default ceiling is a thousand live entries per server, adjustable upward if a community genuinely outgrows it. Because each entry is a structured document rather than a one-line note, a thousand of them holds a great deal more than the number suggests.

The limit that actually binds first is your daily token budget, which is what both answering and teaching draw on:

When the cap is reached the AI stops cleanly and tickets keep working manually. There is no overage bill and nothing breaks. What that budget buys in practice is covered in how much an AI ticket bot costs.

What language is our memory stored in?

Memory is stored in English regardless of the language it was taught in, and replies come back in the language the member wrote in. A policy your team teaches in English answers a question asked in Portuguese, and neither side has to think about it.

That is a bigger deal for mixed communities than it looks, because it means you maintain one set of facts instead of one per language. The wider language story, including the thirty seven languages the bot's own interface is translated into, is in multi-language Discord support.

Can we see and edit what the AI knows?

All of it. The Memory tab in your dashboard lists every entry with its title, full text, tags and type, and gives you four things worth doing about once a month:

The "most used" count is the one people overlook. An entry sitting at zero uses after a month is either a fact nobody asks about, or a fact worded so differently from how members actually ask that it never gets found. Both are useful to know, and neither is visible any other way.

Reading a memory that belongs to a different server is not possible even with its exact identifier. The lookup checks ownership before it returns anything and reports a foreign record as not found. That is a read-time check rather than a filter applied afterwards, which is the difference between separation you can rely on and separation you hope holds.

Why the AI never mentions its own memory

There is a deliberate rule that the AI must never expose the memory system to the person it is talking to. It does not say "brain", "memory", "my knowledge base", "let me search" or "the system". It answers as someone who knows the answer, or admits it does not know and hands off.

This came from a real incident. An earlier version of the instructions used the word "brain" internally, and the AI parroted it back to a Discord member, telling them what "the brain here focuses on". It read exactly as badly as it sounds. Members are talking to your support, not to a retrieval interface, and the fix was to strip the internal vocabulary out entirely.

That rule is also why you cannot judge from a ticket whether an answer came from memory. If you want to know, the Memory tab's usage counts will tell you, and the ticket transcript will show you exactly what was said.

Where per-server memory is the wrong tool

The honest list. If any of these are dealbreakers, this is not the right approach for you, and it is better to find that out now.

Two of those deserve a sentence more. A wrong fact taught confidently stays wrong until a person fixes it, so treat training the way you would treat editing a pinned message. And because nothing re-crawls on a schedule, the most common cause of a stale answer is not the AI at all, it is a policy page that changed three months ago and a training job nobody re-ran.

What happens to our memory if we remove the bot?

Everything tied to your server is purged forty eight hours after the bot leaves. Memory, transcripts, panels, blacklist entries and audit rows. The delay is there so an accidental kick or a permissions mistake can be undone. After it, the data is gone and adding the bot back gives you an empty memory.

Individual members are covered separately. A member can have their data exported or erased across everything the AI holds for them, which is what makes the arrangement workable under GDPR. The details are in our privacy policy.

For the record on where processing happens: messages are processed by our own AI platform (Nexus) and Anthropic's model API. Nothing is sent to other AI providers, and we keep memory on our own infrastructure.

The short version

Per-server memory is a contract, not a mood. The AI knows what is in this ticket, what your staff taught it, and one line naming the ticket. It does not know your pinned messages, your rules channel or your member list, and it will not learn any of them by sitting in your server. What it does know, you can read, search, correct and delete.

That is a smaller claim than "our AI understands your community", and it is the one we can actually stand behind. It is also the easiest thing in this article to check for yourself, on the free plan, in about five minutes: