Diese Seite ist in deiner Sprache noch nicht verfügbar — wir zeigen die englische Version.
One memory. Ten agents. Every single run.
You explain your pricing once. You explain your tone once. You say "the accounting address is compta@…" once. From that moment on, all ten of your agents know it — not because you copied it into ten places, but because they all read the same memory.
That memory is the Neural Core. It's the thing that turns ten chatbots into a team that works for the same company.
The one-line difference
Per-chat memory is a silo: what you told one conversation, the next one cannot see. The Neural Core is a substrate: it sits underneath all ten agents, and none of them can be started without it.
Why "shared" is an architectural fact, not a feature toggle
Most AI memory is a convenience layer bolted on top: a setting you can switch off, a per-chat blob, a "reference previous chats" checkbox. Ours is load-bearing.
- It is mounted unconditionally. The Neural Core sits in the internal provider array that every agent gets — no OAuth token, no flag, no per-agent opt-in. Email is conditional. Documents are conditional. Memory is the floor.
- The agent id is ignored. The function that renders your memory into an agent's context takes an agent id parameter and never uses it. All ten agents receive a byte-identical view of your business memory. There is no per-agent silo to reconcile, because there is no per-agent filtering code.
- Long-term memory is cross-agent by query. The preferences the system learns about you are fetched by your user id only — there is no filter on which agent learned them. Tell Maya "no emojis in emails" and Ashley knows on her next run. The agent that learned it is recorded as provenance, nothing more.
- The team shares a workday. Before every run, each agent is handed a
[TODAY]digest of what already happened across all your conversations and all your agents in the last 18 hours — with the instruction not to re-do finished work. A per-chat memory cannot produce this. A chat cannot see its siblings.
This morning I asked Maya to email the supplier about the delayed order. What's the status, and can you draft a short LinkedIn post about our new lead time?
Ask that of Ashley — who has never touched your inbox and never will — and she still knows what happened this morning, because the workday is shared even though the mailbox isn't.
What goes in
Six categories, each with a real schema. Not a prose blob — typed records.
| Category | What it holds | Why the structure matters |
|---|---|---|
| Brand | Identity, slogan, sector, contact details, socials, light + dark logos, colours, font, tone, voice adjectives | Every agent writes in the same voice and uses the same logo |
| Offers | Product name, real price, currency, description, notes | A price is a price, not a sentence — agents can quote it |
| People | Full name, email, phone, company, role, notes | Maya checks here before asking you for an address |
| Rules | Standing instructions, typed as do / don't / constraint / preference | "Never use emojis in emails" applies to all ten agents |
| Knowledge | Documents, uploaded files, imported pages, images, spreadsheets | Searchable inside the text, not just by title |
| Tasks | What needs doing, with a due date, a priority and an assigned agent | See Tasks |
Because the records are typed rather than prose, an agent can set one cell of a stored spreadsheet by row number and column name, or change one field of a contact. Prose memory can't be edited by cell.
How it gets there
You don't manage a knowledge base. You work, and it fills up.
1. You just say it
No /remember command. No prefix. No keyword. Type "actually our new price is €49" or "le mail de la compta c'est compta@x.com" mid-conversation and it's captured — in whatever language you were already typing in. Behind the scenes an extractor compares what you said against what already exists and decides on its own whether to create, update, merge, or ignore.
If you want to be explicit, say it explicitly:
Remember that our Tuesday newsletter always goes out at 9am, and never with emojis in the subject line.
That creates a real rich-text document you can open and edit later — not a hidden note.
2. You upload a file
Up to 50 MB (20 MB for video). Images, video, PDF, CSV, XLSX, XLS, ODS, TSV, TXT, MD, DOCX, PPTX, JSON, SVG.
- Upload an image and it comes back described — description, mood, key elements, and a category (photo, illustration, screenshot, logo, document, infographic).
- Upload a video and it's described from its frames and categorised (trailer, tutorial, presentation, interview, lifestyle, product, animation).
- Scanned PDFs still work — a PDF is read as a document, not as a text file, so a scan is described and its text extracted like any other. (PDFs above 4 MB are stored but not auto-analysed.)
- Drop a messy CSV on the People or Products card and it's parsed into real contacts or products: the separator is auto-detected, and headers are fuzzy-matched across 10+ languages — including Arabic, Japanese, Chinese and Russian. An import folder is created for you.
3. You paste a URL
Any page. Title, description, favicon, OG image, site name and the extracted text are stored. The fetch is SSRF-hardened: private IP ranges, loopback, link-local and cloud metadata endpoints are all blocked, and the destination is re-validated at every one of a maximum five redirect hops.
4. You connect Shopify
Your products, customers and orders sync in automatically — and your products land locked. See "Your source of truth stays yours" below.
5. Your work fills it in
Send an email through Maya and the recipient is captured as a contact — no LLM guesswork, a deterministic upsert. Generate images in a social draft and you can save them straight into your knowledge, in a folder named from the caption. Like an answer an agent gave you? Hit Save to Neural Core on the message.
How it's used
Before every run — interactive or scheduled — the agent is handed your Neural Core, your standing rules, your known contacts, the preferences learned about you, and the last 18 hours of team activity.
Three details worth knowing:
Facts carry their confidence. Each entry is labelled in the prompt as ✓ confirmed, high confidence, medium confidence or low confidence, so the model knows how much weight to give it. Something you typed yourself outranks something inferred.
Private means private — in the prompt, not just in the UI. Toggle the lock on any entry and it is prefixed, inside every agent's context, with:
[PRIVATE — do not mention in any generated content, emails, or posts]
The agent can still reason with it. It won't put it in your newsletter.
Shared memory makes runs cheaper, not costlier. The volatile blocks are deliberately injected after the static prefix so prompt caching still hits. Memory that grows doesn't mean bills that grow proportionally.
The four promises other memories don't make
Deleted stays deleted
Delete an entry and two things happen: derived facts that came from it are purged, and a tombstone is written. That tombstone is re-injected into the extractor on every future pass under the heading "SUPPRESSED ENTRIES — the user explicitly deleted these. DO NOT RE-EXTRACT."
Systems that quietly re-learn what you just deleted are the norm. This one has a designed defence. Deleting from the app and asking an agent to delete do the same two things: purge the derived facts, write the tombstone.
It won't learn the AI's own hallucinations
This is the failure mode of naive auto-memory: the model invents a plausible detail, the memory extractor reads the transcript, and the invention becomes a "fact" about your business forever.
Our extractor reads your messages in full. It reads the assistant's messages truncated to an 80-character topic hint, with an explicit rule forbidding it from extracting anything an agent said. The comment in the code is blunt: never include full assistant text — it may contain fabricated facts.
Your memory is built from what you said.
Your memory can't be forged
The long-term memory table has SELECT and DELETE policies only — there is no INSERT or UPDATE policy for a logged-in user. Nothing running in a browser can write a memory into your account. Only the server-side extractor can, and you can read and erase every row of it.
You can wipe your memory. Nothing can plant one.
Your source of truth stays yours
- Shopify-synced products are read-only to agents. Every write tool rejects them. Your store is the source of truth; Unyo is a reader.
- Your brand is read-only to agents, enforced across four independent layers — two code guards and two prompt rules. Agents read your logo, colours, tone and voice. Only you change them, in Settings → Brand, where one save writes through to every place that needs it.
Memory costs zero credits
All 23 Neural Core tools are priced at 0 credits — reading, searching, saving, editing, exporting, all of it. Only the model run itself is billed, at real cost. Your memory is not a metered upsell. See Pricing.
You can see all of it
The Neural Core isn't a black box you have to trust. It's a page in the app.
- A hero grid for Products, Contacts and Files, plus Rules, Knowledge and Tasks.
- Seven specialised viewers — rich-text documents, images, video, CSV, PDF, links, structured forms — each editable.
- Nested, coloured folders and multi-select to move a lot of things at once.
- One search box across names, values, summaries, document text, contact emails and phones, and product descriptions.
- Export any category to CSV or JSON, then hand it to an agent to save to Drive or Notion.
- A lock on any entry, and a delete that means it.
Honest limits
We'd rather you find these here than after signing up.
- It doesn't capture everything you say. The extractor has a confidence floor and emits a bounded number of memory items per pass. It catches what matters and it's conservative by design — if something must be remembered, say "remember that…" and it becomes a real document.
- Listing has ceilings. An agent listing contacts returns up to 50; listing products, up to 30; exporting a category, up to 200 entries. The app's own page shows you everything.
- Brand and Shopify products are read-only to agents — by design, but it means "Ashley, change my logo" won't work. You do that in Settings → Brand.
- Steve doesn't author your memory. The website builder is denied the Neural Core write tools — saving, editing documents and spreadsheets, managing rules, creating tasks — at the runtime layer, not just in his prompt.
FAQ
How is this different from ChatGPT's memory?
Two things. First, scope: their memory belongs to a chat or an account and is a feature you can switch off; ours is mounted unconditionally for all ten agents and the code that renders it literally ignores which agent is asking — every agent gets the identical view. Second, structure: theirs is prose, ours is typed records with real fields, which is why an agent can set one cell of your stored spreadsheet or change one field of a contact. And ours is fully visible and editable on a real page, not a list of sentences you can only delete.
Do I have to tell each agent separately?
No — that's the whole point. Tell Maya once that you never want emojis in emails. Ashley reads the same rule on her next run. The preferences learned about you are queried by your account, not by the agent that learned them.
Can an agent see something I marked private?
It can read it — that's what makes it useful for reasoning — but every agent's context carries an explicit instruction not to mention it in any generated email, post or document. If you don't want it in the system at all, delete it: the deletion writes a tombstone that stops it being re-learned.
What happens if the AI makes something up? Does it end up in my memory?
No. The extractor only reads your messages in full. Assistant messages are truncated to an 80-character topic hint and there's an explicit rule against extracting anything the AI said. Memory is built from what you wrote, not from what the model generated.
Can someone inject a fake memory into my account?
The long-term memory table has no INSERT or UPDATE policy for logged-in users — only SELECT and DELETE. No client can write to it. Only the server-side extractor writes, and every row stays yours to read or erase. More on how we enforce this: Security.
Does using the Neural Core cost credits?
No. All 23 Neural Core tools cost 0 credits — reads, searches, saves, edits and exports. You're only billed for the model run itself, at real cost. See Pricing.
What if I delete my account?
Everything goes. Account deletion purges the whole record atomically and revokes your tokens at every provider that offers a revoke endpoint. See Your data and deletion.
Which languages does it understand?
You don't need a command or a prefix — state a fact in the language you're already typing in and it's captured. Imported spreadsheet headers are fuzzy-matched across 10+ languages, including Arabic, Japanese, Chinese and Russian. Agents reply in the language you choose, independently of what you type.
Start with one fact
Open any agent and tell it something true about your business. Then open a different agent and ask about it.
That's the demo.
Meet the ten agents → · How your data is protected → · Pricing →