Reference memory for AI creative agents

Turn any public video into a reference your agent can cite.

Give Refmatter a URL. Get back a verified reference: the source as it was observed, the media itself with a poster frame, and provenance that says how and when it was acquired. Through a REST API, a TypeScript SDK, and MCP.

Read the quickstart Request early access

resolve.ts@refmatter/sdk

// npm install @refmatter/sdk
import { Refmatter } from '@refmatter/sdk';

const refmatter = new Refmatter({ apiKey });
const { reference } = await refmatter.resolve(
  'https://www.youtube.com/watch?v=jNQXAC9IVRw',
);

referenceapplication/json

{
  "state": "ready",
  "platform": "youtube",
  "sourceObserved": { "title": "Me at the zoo", "durationSeconds": 19, … },
  "machineDerived": { "container": "mp4", "width": 320, "height": 240, … },
  "provenance": { "contentHash": "sha256:…", "acquiredAt": "2026-09-14T14:21:49Z" },
  "mediaAvailability": "ready",
  "mediaHold": { "expiresAt": "2026-10-14T14:21:49Z" },
  …
}

What comes back

One URL in. One reference out, with everything a creative agent needs to trust it.

ObservedThe source, as it was: Title, description, channel, publish time, and counts exactly as the source served them at acquisition. Missing values stay null; nothing is guessed.
MediaVerified bytes and a poster: The video itself, transferred through our own pipeline, decoded to prove it plays, with a poster frame. 480p analysis copies by default, 1080p when you ask.
ProvenanceEvidence you can cite: Content hash, capture timestamp, source route and contract version travel with every reference, so a generator’s input can always be traced back.
YoursNotes, tags, a title: Human-authored fields that a refresh never overwrites, written with optimistic concurrency so two tools cannot clobber each other.

Built for agents

Three ways in. One contract underneath.

REST API

Idempotent imports, a cursor-paged catalog, media holds, and download URLs. Every route is described in the OpenAPI document served by the API itself.

GET api.refmatter.com/openapi.jsonAPI reference ↗

TypeScript SDK

Generated from the contract, so it cannot drift. One call from URL to reference, typed errors with a retryable flag, the raw client when you need it.

npm install @refmatter/sdkSDK guide ↗

MCP server

The same operations as tools for Claude, Cursor, or your own agent. Each client brings its own workspace key; the server holds none.

claude mcp add --transport http refmatter …/mcpConnect an agent ↗

Sources

Honest about what a source will give.

Only public, unauthenticated content, acquired through our own tooling. No accounts, cookies, client emulation, or rotating proxy pools. When a source will not serve something publicly, you get a stable failure code, not a half-empty result.

YouTubePublic videos and ShortsEarly access
VimeoPublic videosPlanned
TikTokPublic videos and profilesPlanned
InstagramPublic post or reel by URLPlanned
Reddit, PinterestPosts, pins, boardsPlanned
Ad librariesMeta, TikTok, GooglePlanned

Retention you control

Derivatives stay. Bytes are time-boxed.

Metadata, poster, provenance, and your notes live as long as the reference. The video bytes exist only under a media hold you own: 30 days by default, extendable to 365, releasable any time. No standing archive of other people’s video, and a takedown path that acts within 24 hours.

Read the retention and takedown policy

Kept with the reference: Observed metadata, poster frame, provenance, machine-derived facts, annotations
Held while you say so: Verified video bytes, 30–365 days, then deleted and reported as expired
Never built: A permanent library of third-party media

Early access

Building with the first creative teams.

Refmatter is in early access. YouTube is live; the next sources follow the teams that ask for them. Tell us what your agents need to remember.

Request access Browse the documentation