Corpus: Shared Memory That Learns Your Business, One Signed Entry at a Time

Every decision an agent makes rests on context: what the refund policy is, when the payout batch cuts off, which contractor is approved, what a resident was promised last week. Most teams give agents that context one of two ways. They crawl the file shares and wikis into a vector store, or they bolt on a memory database that any agent can write to and any agent can read.

Both work until the first time something goes wrong. Then the questions start. Who wrote this entry? Is it still true? Has anyone changed it since? Did the agent that acted on it have any business reading it? A plaintext author field and a timestamp don't answer any of that, and by the time you're asking, you're already explaining yourself to a customer or an auditor.

What Corpus is

Vorim Corpus is the memory layer of the Multi-Agent Fabric, and it takes a different position from the start. Knowledge isn't scraped into it; it's written into it by your agents as they do real work. Each entry is attributed to a verified agent identity and signed at the moment it's written, so what your organisation knows is traceable to a source and tamper-evident from day one. A forged or altered entry fails verification. A low-quality write is always traceable to the agent that made it.

Reads are permission-scoped. An agent holding the read scope sees the slice of context it's cleared to see, not an open pool. And there's a human read path too: the Corpus view in the dashboard lets an operator browse everything agents have written, search it, and see at a glance which entries are signed and which are merely attributed.

The SDKs expose it as four calls in TypeScript and Python: write, read, list and delete. If you already give your agents identity and permissions through Vorim, Corpus is the same key and the same scopes.

Why we say it learns your business

Because it fills up the way a good colleague's head does: from doing the job. An agent that reconciles refunds learns that anything over $500 needs a named approver. A payouts agent learns that Friday batches after 15:30 settle on Tuesday. A maintenance agent learns which contractors are approved and up to what amount. Written down at the moment it mattered, by the agent that found it out, and signed.

That accumulation is what turns a memory store into working knowledge. A new agent, or a new hire looking at the dashboard, starts on day one with what the organisation actually knows about how it runs, and can see who found each thing out and when.

Live context before the decision

Corpus is read at the moment an agent acts, not from last night's export. So the decision is made with current context, and the permission check on the action that follows is made against the agent's own scopes, as every Vorim action is. The action gets signed into the audit record the way it always has.

Where it sits

Corpus is not a replacement for your vector store or your orchestrator's memory. It's the governed layer underneath them for the knowledge that has to be accountable: the rules, thresholds and commitments an agent acts on with real money and real records. Keep the tools you use. Put the knowledge that matters somewhere you can answer for.

If your agents are working from context nobody can vouch for, that's exactly what we build.

Book a demo →