neural-core-knowledge
How your agents learn facts about your business — say "remember that…", let the extractor pick things up on its own, mark entries private, set standing rules, and delete what you don't want kept. Knowledge is shared by all 10 agents and costs no credits.
Diese Seite ist in deiner Sprache noch nicht verfügbar — wir zeigen die englische Version.
Every fact your agents know about you lives in one place: the Neural Core. Not one memory per chat, not one memory per agent — one memory, read by all 10 agents on every single run.
Tell Maya "no emojis in emails" and Ashley writes without emojis on her next post. That isn't a setting you switch on. It's how the system is built: the Neural Core is loaded into every agent's context before every run, with no agent filter, no category filter and no on/off toggle.
You'll find it in the sidebar under Neural Core. This page covers knowledge entries, rules and memory — the things your agents learn. Contacts, products and files have their own pages: Contacts · Products · Files.
Memory is free
All 23 Neural Core tools cost 0 credits. Reading, searching, remembering, editing, exporting — none of it is metered. You're only ever billed for the agent run itself, at its real cost. Memory is not an upsell.
Three ways a fact gets in
| How | What you do | What you get |
|---|---|---|
| You ask for it | "Remember that…" | A real rich-text document you can open and edit |
| It picks it up | Just talk normally | A structured entry, extracted in the background |
| You type it yourself | Add an entry in the Neural Core | Full control, marked "✓ confirmed" |
1. Say "remember that…"
The explicit route. Works with any of the 10 agents, in the language you're already speaking.
Remember that our refund window is 30 days, no questions asked, but customers have to email support before shipping anything back.
The agent creates a rich-text document in your Knowledge category — the same format as if you'd written it in the Neural Core editor yourself. It gets a title, a body you can format, and it opens in the document editor when you click it. It is not a scrap of text buried in a log.
The trigger phrase doesn't matter much — "remember that…", "don't forget…", "note this down…" all work, in any language. What matters is that you're asking the agent to keep something.
Steps:
- Open a conversation with any agent.
- Type your fact, prefixed with remember that… (or your language's equivalent).
- The agent confirms in one short sentence — it won't dump the raw entry back at you.
- Open Neural Core → Knowledge to see it. Click to edit, format, rename or move it into a folder.
Ask an agent to edit it later
You don't have to open the app. "Rewrite the refund note to say 14 days" or "add a paragraph about exchanges to the refund note" works — the agent reads the full document first, edits it, and keeps your formatting intact.
2. Automatic extraction — no command, no prefix
This is the part people don't expect. You never have to tell an agent to remember anything.
Type a price change, an email address, a constraint, a new product, a person's role — mid-sentence, in the middle of a normal request — and a background pass picks it up and files it in the right category, in whatever language you wrote it in.
You asked for a draft. You got a draft. You also now have a price change in your Neural Core, and every other agent will quote €49 from that moment on.
The extractor doesn't just append. Each fact it finds carries a decision — create, update, merge or ignore — checked against what you already have. Change a price twice and you get one entry with the right number, not two contradicting each other.
Six categories it writes into: audience, offers, goals, people, rules, knowledge.
It will not learn the AI's own words
The extractor sees your messages in full. It sees the agent's replies truncated to an 80-character topic hint, and it's explicitly forbidden from extracting facts from them. This is deliberate: an AI that learns from its own output eventually believes its own mistakes. The comment in the code is blunt about it — assistant text may contain fabricated facts, so it never reaches the extractor. Your Neural Core learns from you, not from the model.
It doesn't catch everything — on purpose
There's a confidence floor: a fact the extractor isn't reasonably sure about is dropped rather than guessed. It also caps how much it takes from any single exchange. "It remembers everything" would be a lie. If a fact matters, say "remember that…" — that route is explicit and always creates an entry.
3. Confidence labels — your agents know how much to trust a fact
Not every fact is equal, and your agents are told so. Every entry carries a label into the prompt:
| Label | What it means |
|---|---|
| ✓ confirmed | You stated it explicitly, or typed it yourself |
| high confidence | The extractor was very sure |
| medium confidence | Reasonably sure |
| low confidence | Kept, but flagged as soft |
You'll see the source of each entry in the Neural Core too — whether you added it or an agent did — and you can filter by it. If something's wrong, edit it: an entry you correct by hand becomes ✓ confirmed, the strongest signal there is.
4. Private entries — the lock is real
Click the lock on any entry and it becomes private.
A private entry is still read by your agents — they still know it, still use it to reason. But the entry arrives in their context wearing a prefix:
[PRIVATE — do not mention in any generated content, emails, or posts]
For all 10 agents, simultaneously. So the agent can know your margin is 62% and still never put it in a customer email.
What to lock
Margins and cost prices. Internal notes on a client. Anything true that you'd never want appearing verbatim in a draft. Lock it — don't delete it. A locked fact still makes your agents smarter; a deleted one doesn't.
A lock is an instruction, not a firewall
The private prefix tells the model not to repeat the fact in generated content. It's a strong, prompt-level control applied to every agent at once — but it is not the same thing as the fact never being loaded. If something must never enter a model's context at all, don't put it in the Neural Core.
5. Rules — say it once, every agent obeys
Rules are their own category: standing instructions that shape everything your agents produce.
Ask any agent to list, add, update or delete your rules in chat, or manage them in the Neural Core. Each rule is typed — a do, a don't, a constraint or a preference — so the model knows whether you're expressing a hard limit or a taste.
The point: you set it in a conversation with Maya, and Lucy, Ashley, Blake and the rest follow it on their next run. There is no per-agent settings screen to keep in sync, because there is no per-agent memory to keep in sync.
6. Delete means deleted — and it stays deleted
The failure mode of every auto-memory system: you delete a fact, then mention it in passing, and the extractor cheerfully re-learns it. You end up fighting your own tool.
Unyo closes that loop. When you delete an entry — in the app or by asking an agent — two things happen:
- The entry is removed.
- A tombstone is written recording what was there.
That tombstone is injected back into the extractor on every later pass, under the heading SUPPRESSED ENTRIES — the user explicitly deleted these — DO NOT RE-EXTRACT. The instruction is absolute: even if the same information turns up in a later conversation, the extractor is told to ignore it.
Deleting a file or a link also purges the facts that were derived from it — not just the file card.
Deletion is permanent
There's no undo and no version history on Neural Core entries. Deleting through an agent is confirmation-gated, and an agent that's unsure which entry you mean will list the matches and ask once before removing anything. But once it's gone, it's gone.
7. Reading it back
Ask any of the 10 agents what they know. Search runs across entry names, values and summaries — and inside your content: document text, contact names, emails and notes, product descriptions, extracted file text.
- Search returns matches with a short preview.
- Read the full entry returns the complete document text (up to 15,000 characters) or the whole spreadsheet with its headers.
- Export a whole category as CSV or JSON — up to 200 entries — and have the agent save it straight to Drive or Notion.
In the app, one search box covers everything, and you can sort by recent or oldest and filter by whether you or an agent created the entry.
8. A shared workday, not just shared facts
Beyond the entries you can see, agents carry a short-term working memory: things you've told them before, and a rolling record of what already happened today across all your conversations, with every agent, for the last 18 hours.
So you can ask Blake about something you did with Maya this morning, and he knows. And no agent re-does work another one already finished.
This memory can't be forged
That conversational memory can only be written by the extraction service running server-side. There is deliberately no way to insert or edit it from a browser — no client, no script, nothing signed in as you can inject a memory. It's readable and erasable by you, and writable by nothing else.
What agents can't do to your Neural Core
The boundaries are enforced in code, not asked for politely in a prompt:
- Brand is read-only to every agent. Your identity, slogan, colours, logos, tone and voice adjectives are yours to set in Settings → Brand — agents read them, and are refused if they try to write them. Four independent layers enforce this. See Brand Kit.
- Shopify-synced products are locked. If Shopify is your source of truth for a product, no agent can quietly overwrite it —
brain.update,brain.deleteand the product write tools all reject it. - The extractor never writes Brand, even if you describe your brand in a conversation. It's rejected in the prompt and in the code.
Tips
- Correct, don't re-explain. "Actually the price is €49" is enough. The extractor will find the existing entry and update it rather than creating a duplicate.
- Lock before you launch. Do one pass over your Neural Core and lock anything you'd hate to see in a customer-facing draft. It takes two minutes and it's the single highest-value thing on this page.
- Use "remember that…" for anything load-bearing. Automatic extraction is a bonus, not a guarantee. Prices, policies and hard constraints deserve the explicit route.
- Write rules as rules. "Never use emojis" beats "I'd prefer fewer emojis I think" — a typed rule is a stronger instruction than a vague preference.
- Folders are for you, not the agents. Agents read everything regardless of folder. Organise for your own sanity.
Troubleshooting
"I told the agent something and it didn't get saved." Extraction has a confidence floor and skips very short acknowledgements. Say "remember that…" explicitly — that route always creates an entry.
"An agent quoted an old price." Check the Neural Core for two entries saying different things. Delete the wrong one — the tombstone will stop it coming back — and edit the right one so it becomes ✓ confirmed.
"A fact I deleted came back." It shouldn't; deletion writes a tombstone that blocks re-extraction. If it reappeared, the likeliest cause is a different entry saying the same thing, or a file/link you imported that still contains it. Search for the value, not the entry name.
"An agent said it can't edit my brand." That's correct and deliberate. Brand lives in Settings → Brand and only you can change it.
"An agent said a product is locked." It's synced from Shopify. Change it in Shopify; the Neural Core follows.
FAQ
Do I have to tell agents to remember things?
No. Facts you mention in normal conversation — prices, email addresses, constraints, people, products — are picked up in the background, in any language, with no command or prefix. But extraction has a confidence floor and takes only a limited number of facts per exchange, so anything load-bearing is worth saying explicitly with "remember that…". That route always creates an entry.
Does every agent see the same memory?
Yes — identically, and not by convention. The Neural Core is loaded into every agent's context on every run, filtered by your account and nothing else. There is no per-agent view, no per-agent filter, and no memory toggle. Tell one agent something and the other nine know it on their next run.
If I mark an entry private, can agents still use it?
Yes — that's the point. Private entries are still read by all 10 agents, but they arrive tagged "do not mention in any generated content, emails, or posts". So an agent can reason with your margin and still never write it into a customer email. If a fact should never reach a model at all, don't store it in the Neural Core.
If I delete something, will it be re-learned from an old conversation?
No. Deleting — through the app or by asking an agent — writes a tombstone that is fed back into the extractor as "the user explicitly deleted this — do not re-extract", even if the same information appears in a later conversation. Deleting a file or link also purges the facts derived from it. Deletion is permanent: there's no undo and no version history.
Can the AI's own mistakes end up in my memory?
Not from automatic extraction. The extractor reads your messages in full but only ever sees an 80-character topic hint of the agent's replies, and it's explicitly forbidden from extracting facts from them. An agent's own words never become memory through automatic extraction.
Does memory cost credits?
No. All 23 Neural Core tools are priced at zero credits — reading, searching, remembering, editing, exporting. You're billed for the agent run itself at its real cost, nothing more. See Credits & billing.