Neural Core

Most AI tools give you a memory per chat. Close the tab, open a new one, start over. Use a different assistant, explain yourself again.

Unyo has ten agents and one memory. It's called the Neural Core, it lives at /brain, and every agent reads it on every single run - Ashley when she writes a post, Maya when she drafts a reply, Blake when he analyses a file, Steve when he builds a page. Tell Maya your price went up to €49 and Ashley won't quote €39 in tomorrow's caption.

This isn't a convention the agents politely follow. It's how the system is wired: the Neural Core is mounted for every agent with no toggle and no setting, and the function that renders your memory into an agent's context ignores which agent is asking. All ten get a byte-identical view.

The six categories

Your Neural Core isn't one long prose blob. It's six typed categories, each with a real schema - which is exactly why an agent can set one cell of a spreadsheet or store a real price with a real currency, instead of rewriting a paragraph and hoping.

Four ways facts get in

1. Just talk to an agent

After you send a message, a background pass reads it and decides what's worth keeping - create a new entry, update an existing one, merge it, or ignore it. It works out that "actually make it 49" refers to the product you mentioned two messages ago.

Two honest limits worth knowing:

• It only keeps facts it's reasonably confident about, and it captures a handful of items per pass - not every sentence you type.
• It reads your messages in full, but the agent's replies are deliberately reduced to a short topic hint before extraction, and the extractor is explicitly forbidden from learning anything the AI itself said.

That second one matters more than it sounds. The classic failure of automatic memory is an assistant inventing a detail, remembering its own invention, and treating it as fact forever. Unyo closes that loop on purpose: your Neural Core learns from you, not from the model.

2. Ask for it explicitly

This creates a real rich-text document in Knowledge - one you can open and edit in the Neural Core, not a hidden note.

3. Upload a file

Drag a file onto a category, or use the uploader.

• Up to 50 MB per file - 20 MB for video.
• Images, video (MP4, MOV, WebM), PDF, CSV, XLSX, XLS, ODS, TSV, TXT, MD, DOCX, PPTX, JSON, SVG.
• Upload an image and it's described automatically - description, mood, key elements, and a category (photo, illustration, screenshot, logo, document, infographic).
• Upload a video and you get a description built from its frames, categorised (trailer, tutorial, presentation, interview, lifestyle, product, animation).
• Drop a messy CSV or a PDF on People or Offers and it's parsed into real contacts or products. The separator is detected for you, and headers are fuzzy-matched across more than ten languages - including Arabic, Japanese, Chinese and Russian. An import folder is created automatically.

4. Paste a URL

Paste any link and the Neural Core fetches it and stores the title, description, favicon, preview image, site name and the extracted text. The fetcher refuses private IPs, loopback, link-local and cloud metadata addresses, and re-checks at every redirect hop - so a link someone sent you can't be used to make Unyo read something it shouldn't.

You can also click Save to Neural Core on any agent message to keep what it said as durable memory - that one is your call, explicitly, which is the point.

What the /brain screen gives you

Everything the agents read, you can see and change.

• A three-card hero - Products, Contacts and Files, front and centre.
• Seven specialised viewers, chosen automatically from what the entry actually is: rich-text document editor, image viewer, video viewer, CSV/spreadsheet editor, PDF viewer, link viewer, and a structured form for typed entries like contacts and products.
• Nested, colour-coded folders - organise Knowledge the way you'd organise a drive.
• Multi-select - tick several entries and move them at once.
• One search box for everything - it looks inside names, values, summaries, document text, contact emails and phone numbers, and product descriptions. Not just titles.
• Filters and sort - newest or oldest, and filter by who created an entry: you or an agent.
• The private lock 🔒 - see below.
• Download what you open - a stored spreadsheet exports as CSV, a document as HTML, and files download as themselves. To pull a whole category out as CSV or JSON, ask an agent (see below).

The private lock

Flip any entry to Private and no agent will mention it in a generated email, post or document.

This is not a UI decoration. A private entry is still visible to the agent - it can use the fact to reason - but it arrives in the agent's context wrapped in an explicit instruction not to reproduce it in anything it generates. For all ten agents, at once.

Delete means deleted - and it stays deleted

When you delete an entry, two things happen. The row goes. And a tombstone is written that gets handed to the extractor the next time it runs, under the heading "the user explicitly deleted these - do not re-extract".

That's the difference between deleting a fact and deleting it for good. Systems that quietly re-learn what you just removed - because the fact is still sitting in an old conversation - are the norm. This one is built to refuse.

Deleting an entry that produced other facts (a file, an imported link) removes the facts derived from it too. And it works the same way whether you delete it yourself or ask an agent to.

Ask any agent about your own memory

You don't have to open /brain to use it.

Agents can search inside document text, read a whole stored document or spreadsheet back to you, edit a document (rewrite it, append to it, or find-and-replace an exact passage while keeping the formatting), edit a spreadsheet (replace the data, append rows, or set one cell by row number and column name), add or update contacts and products, and manage your standing rules.

Why one shared memory beats ten good ones

Preferences travel. Tell Maya "keep replies short, no emojis" and Ashley honours it on her next run. Your long-term memory is looked up by you, not by which agent is talking - there is no per-agent filter in the query at all. The agent that learned a thing is recorded as provenance, and nothing more.

The team shares a workday. Every agent also gets a short brief of what already happened across all your conversations, with all your agents, in the last 18 hours - so you can say "finish what Blake started this morning" and be understood, and so two agents don't redo the same work. A per-chat memory can never do this: a chat cannot see its siblings.

Facts arrive with their confidence. Each entry reaches the agent labelled ✓ confirmed, high, medium or low confidence. The model is told how much to trust what it's reading, rather than being handed a flat wall of assertions.

Your sources of truth stay yours. Products synced from Shopify are locked - every agent write tool refuses them, so nothing can quietly overwrite your store. Brand is read-only to agents, enforced in four independent places: you manage it in Settings → Brand, and the agents live with it.

Your memory can't be forged. Your long-term memory table can be read and erased by you, and written by nothing on the client side - only the background extractor writes it. You own the erase button; no page, no script and no agent has an inject button.

Set your Brand once

Settings → Brand is where identity, slogan, sector, contact details, socials, light and dark logos, colours, font, tone and voice adjectives live. Fill it in once and all ten agents use it - the same logo in a design, the same colours in a site, the same tone in an email.

It saves as you type, and it lands everywhere it needs to at once: your Neural Core, your branding settings, your company name and the sidebar - one action, no "did it save?".

Tips

• Front-load the boring stuff. Ten minutes filling in Brand, your three main Offers and your top Rules will improve every output from every agent, permanently. It's the highest-leverage ten minutes in the product.
• Write rules as rules, not as requests. "Never use emojis in emails" is a rule. "Can you avoid emojis this time?" is a message.
• Use folders for Knowledge. Once you've imported a few files and links, colour-coded nested folders are the difference between a library and a pile.
• Lock before you ask. If a fact should never appear in an output, set it Private before you ask for the output, not after.
• Check what it learned. The Neural Core shows you what came from you and what came from an agent. Filter by source now and then - correcting one wrong entry is faster than correcting every email that used it.

Troubleshooting

"I told an agent something and it's not in the Neural Core." Not every sentence becomes an entry - short acknowledgements are skipped, and only facts the extractor is confident about are kept. If it matters, say it explicitly: "Remember that our refund window is 14 days." That creates a real entry every time.

"An agent didn't list all my contacts." Listing is capped: an agent can list up to 50 contacts and up to 30 products in one go. That's a hard cap, not a page-through. For the full picture, open /brain - or ask for an export, which goes up to 200 entries per category.

"A contact or product looks empty in the card." Open it in /brain and check the fields directly. If it was created from a messy import, the header mapping may have missed a column - edit it in place, and it's fixed for every agent at once.

"An agent won't change my product." If it's synced from Shopify it's locked on purpose, and every agent write tool refuses it. Change it in Shopify and let it sync.

"An agent won't update my brand." By design. Brand is read-only to agents. Edit it in Settings → Brand.

"My file won't upload." Check the ceilings: 50 MB for files, 20 MB for video. For People and Offers, only importable formats are accepted (PDF, CSV, XLSX, XLS, ODS, TSV, TXT, MD, DOCX) - an image dropped on Contacts has nothing to import.

Next: Knowledge · Contacts · Products · Files · Brand Kit · Tasks