Agents that remember what they promised

Hyperstruck is the intelligence layer under your agents. When a run finishes, two things stay. Learnings are what to do next time. Facts are what is true about a customer, invoice, or system. This cycle, agents remember what they promised and when it is due.

Agents remember what they promised

Agents can record what they owe, due in the reader's timezone. A note becomes a dated commitment. You can close it when the work is done.

Example: a sales rep writes "follow up with Acme on Friday." That commitment shows up due Friday for whoever opens it in London or Sydney.

Manage remembered facts in the web app console

You can view, keep, or delete an agent's facts in the web app console. You can mark a source as trusted. You can clear a whole set of facts without a database login.

Example: a platform lead sees two payment terms for Customer X, keeps both, or clears a bad import. Finance marks the ERP export as trusted so those terms are used right away.

Every item on the list, and a count you can show

Agents process every item of a list and report how much they covered. You do not get a "done" on the first twelve.

Example: a finance agent given 50 invoices works all 50 and reports the count. Missed invoices stay visible.

Hand us the customer file. The facts land.

You can hand Hyperstruck a customer file or a policy document. The facts land dated and sourced. A long document is read in full. People named in the notes stay attached to the facts about them.

Example: a 50-row customer export, or a full MSA, lands as facts the next collections or legal agent can use. Jane at Acme stays attached to Jane.

Ask about three customers. Get three records.

Ask about three customers and the agent gets facts for all three, not just the first. A long legal name still finds the record.

Example: "What is on the books for Acme Holdings Limited and Brightside Pty Ltd?" returns facts for both.

A written policy beats a one-off comment

A rule from a signed company policy beats a casual comment. The agent can tell which is which.

Example: the signed collections policy is the default. "Do not chase Customer X this week" stays a one-off, and the agent can see that it is a one-off.

A renamed tool does not erase the learning

A learning earned against one tool still reaches an agent whose tools are named differently. You do not re-earn the same learning because the platform team renamed a function.

Example: a collections rule learned against lookup_invoice still applies when the same job is now called get_invoice. We wrote this up.

Agents learn when not to act

An agent can learn when not to act, from work that was halted before it finished. A tool can refuse to send if the field it needs was not read on this run.

Example: a payouts run stopped because the sort code was unread teaches the next run not to send until that field is confirmed.

Credit for LangGraph, MCP, and the editor

LangGraph, MCP, and editor agents now earn credit when they reuse a rule they were given. You can mint a key that reads only your tenant's remembered facts. Your own MCP server credential is accepted when you connect it. A long coding session gets a fresh set of learnings part-way through, not only at the start.

Example: a LangGraph collections agent that reuses the net-60 rule for Customer X earns credit for that reuse. A finance integration can read those terms with a tenant key, without a database login.

You can see when two facts disagree

Two sources for the same fact stay as two sources. Search shows which learnings contradict each other.

Example: finance's net-60 policy and a one-off exception both stay visible. A platform lead searching "net-30" can see they disagree before the next collections cycle.

Talk to us

Want to see this on your own workflow? Write hello@hyperstruck.com.