Files in your Neural Core

Upload a file once. Every agent can find it, read it and use it — forever.

Your Neural Core is the shared memory all 10 agents read on every single run. Files are the part of it you can see: a product shot, a price list, a brand guide, a demo video. When you upload one, Unyo doesn't just store it — for images, video, PDFs, spreadsheets and text files it describes them, extracts what's inside, and files them where an agent will actually find them. (DOCX and PPTX upload and store, but aren't described or read yet — export to PDF if you want an agent to search inside.)

What you can upload

The ceiling is 50 MB per file, except video which is capped at 20 MB. Both limits are checked before the upload starts, so you find out immediately — not after a three-minute wait.

Upload a file

1. Open the Neural Core from the sidebar.
2. Open the Files card — or any category and folder you want the file to land in.
3. Drag the file in, or click to browse.
4. Wait for the description to appear. Unyo analyses the file in the background; the entry shows up right away and fills in as the analysis lands.
5. Open the entry to check what was understood — and edit it if you want. Every generated description is yours to overwrite.

What Unyo does with an image

Upload a photo and the model looks at it and writes back four things:

• A description — what's actually in the frame.
• A mood — the feel of the image, which is what makes it usable as a style reference later.
• Key elements — the objects, people and details worth matching on.
• A category — one of: photo, illustration, screenshot, logo, document, infographic, or other.

That's why "find me the picture with the wooden counter" works later. The search runs against the description, not the filename.

What Unyo does with a video

Video gets the same treatment, from its frames: Unyo pulls the first and middle frame, sends them to the model together, and stores a description plus a category — one of trailer, tutorial, presentation, interview, lifestyle, product, animation or other.

What Unyo does with documents and data

• PDF — the model reads the document and stores its text with the entry, so agents can search inside it. Scanned PDFs work too. Automatic analysis covers PDFs up to 4 MB; a bigger PDF is still stored, but isn't read.
• CSV / TSV / XLSX / XLS / ODS — parsed into a real table with headers and a row count, and opened in a spreadsheet editor. Not a screenshot of data — data.
• JSON, TXT, Markdown — extracted directly.

Once the text is extracted, it's searchable. Ask any agent "what do you know about our refund window?" and the search runs across entry names, values, summaries and the extracted text of your documents, your contacts' names, emails and notes, and product descriptions.

An agent can also read a whole document or spreadsheet back to you — up to 15 000 characters, plus the headers and row count for a sheet. Past that it's truncated, and the agent is told how much it didn't see.

Smart import: a messy CSV becomes real contacts

Drop a contact list or a price list on the Contacts or Products card and Unyo doesn't file it as an attachment — it parses it into real entries.

1. Open the Neural Core and go to Contacts or Products.
2. Drop in a CSV, TSV, spreadsheet, PDF or DOCX.
3. Unyo auto-detects the separator (semicolon, comma or tab — it counts them across the first lines and picks the winner), so you don't have to care how the file was exported.
4. Headers are fuzzy-matched across 10+ languages — French, English, Spanish, German, Turkish, Portuguese, Italian, plus Arabic, Japanese, Chinese and Russian. Nom, Full name, 姓名 and الاسم الكامل all land in the same column. Separate first-name/last-name columns are recombined automatically.
5. A dedicated import folder is created for the batch, so a bad import is one folder to delete — not 200 rows to hunt down.
6. If the file isn't a clean CSV, the model reads it and extracts the entries anyway.

Import a link

Paste a URL and the Neural Core fetches the page and stores its title, description, favicon, OG image, site name and extracted text — a real entry you can search, not a bookmark.

1. Open the Neural Core.
2. Choose Import a link.
3. Paste the URL.

Seven ways to look at your stuff

The Neural Core picks the right viewer for what you stored, automatically — there's nothing to choose:

Folders, colours and multi-select

• Folders nest — a folder inside a folder inside a folder.
• Each one takes a colour: blue, green, yellow, orange, red, pink, purple or brown.
• Multi-select lets you move many entries at once instead of dragging them one by one.
• Imports get their own folder automatically.

Lock a file to private

Every entry has a Private toggle, shown as a lock. It is not a UI decoration.

Flip it, and that entry is handed to the model with this in front of it: do not mention in any generated content, emails, or posts. Not for one agent — for all 10, simultaneously, on every run. The agent can still know the fact; it will not put it in an email or a post.

Use it for internal margins, supplier names, the notes you'd never want appearing in a caption.

Save an agent's work back into your files

• Save to Neural Core — any agent message has a save button. What it said becomes durable memory.
• Images from a post draft — ask the agent to save the images it just generated into your Neural Core and a folder is created automatically, named from the caption. The agent doesn't hand over URLs; the images are pulled straight from the draft.

Deleting a file

Deleting is permanent, and it's thorough:

• The file entry goes.
• Any facts derived from it go with it — the description, the extracted entries. No orphans left behind quoting a file that no longer exists.
• A tombstone is written, and it's fed back into the system that learns from your conversations as "the user deleted this — never extract it again".

That last one matters. The normal failure mode of automatic memory is that it cheerfully re-learns tomorrow exactly what you deleted today. This one is designed not to.

Who can touch your files

All 10 agents read the same files. That isn't a setting — it's how the product is built: there's no per-agent memory silo and no "memory on" toggle to forget.

Steve, the website builder, is the exception on writes. He can read your Neural Core, and he's blocked at the runtime layer — not by a prompt asking nicely — from saving, editing or deleting your knowledge entries, documents, spreadsheets and rules. He'll use your logo; he won't rewrite your brand guide.

Shopify-synced products are locked. If you've connected Shopify, your synced products carry a lock and every write tool refuses them. Shopify stays the source of truth; Unyo won't quietly diverge from it.

Tips

• Describe once, reuse forever. A good image description is what makes "use the one with the warm lighting" work six weeks later. Edit the generated description if it missed the point.
• Use folders per campaign or per client, and colour them. Search is fast, but folders are how you delete a whole batch cleanly.
• Drop the price list on Products, not Files. Same file, completely different outcome: one gets stored, one gets parsed into real products with real prices.
• Lock the internal stuff on day one. It's a one-click toggle and it applies to all 10 agents at once.
• Trim video before uploading. 20 MB goes fast.

Troubleshooting

The upload was rejected. Check the size first — 50 MB, or 20 MB if it's video — then the format. If you dropped it on Contacts or Products, that card only takes PDF, CSV, TSV, XLSX, XLS, ODS, TXT, MD and DOCX.

The description never appeared. The entry is saved regardless; analysis runs separately in the background. Reopen the entry — and if it's still blank, delete and re-upload.

The PDF has no description or extracted text. PDFs are analysed whole, but only up to 4 MB — a larger PDF still uploads and stores, it just doesn't get read. Split it or compress it and re-upload.

The import created garbage entries. Delete the import folder — the whole batch lives in one folder for exactly this reason — fix the source file and drop it again.

An agent can't find an old image. See the FAQ below — image search has a real ceiling.

Next: Knowledge · Contacts · Products · Brand kit