Welcome to Hyperstruck Docs
Learnings, claims, obligations and the APIs for building on Hyperstruck.
Hyperstruck learns from your files and from how the work went, then hands it to whoever needs it next. Integrate it with LangGraph, connect any MCP host, or work directly from Claude Code.
Recommended reading order
Start with the Learning loop, then connect through the MCP server, and read Memory before diving into the API and integration pages.
Better than searching your documents
Most agent memory is RAG (retrieval-augmented generation): your documents are cut into passages, and the ones that look most like the question are pasted in front of the model. That finds text. It does not know which passage is out of date, it does not notice when two sources disagree, and it learns nothing from whether the work went well.
Hyperstruck keeps what your agent actually needs. Claims hold what is true now about every person, company and thing it deals with, each dated and sourced, with old values ended when a newer source says they changed, and conflicts flagged for a person. Learnings hold what works, earned from real outcomes, gaining standing when they help and losing it when nothing confirms them. Your agent gets the right ones before every task, so each new task starts from what worked last time.
How it works
Memory
The two things an agent remembers: learnings about what to do, and claims about what is true.
Learning
What a learning is, how it makes your agents better over time, and what it is not.
Claims
The specific details your agent knows about people, companies and things, kept up to date with their sources.
Obligations
Claims about what someone owes and by when, open until they are done.
Documents
Send what your team writes and your agents answer from it, pointing to the passage behind each answer.
Agent value
A running total of what each agent is worth, in dollars and time saved.
Insights, Workflows and Evals
Once agents have learnings and claims, they can notice what to do, do a unit of work, and judge the answer.
Insights
Proactive: ranked findings of what is worth doing now, before anyone has to ask.
Workflows
Do the work: first-line support, triage, candidate matching, using knowledge you already hold.
Evals
Judge real workflow runs, then improve that judgment on your data before you roll it out.
How it all fits together
Documents, claims, learnings, agents, Insights, Workflows and Evals, and how each run feeds the next.
Use it
Learning loop
Resolve before the work and observe after it, so the agents you already run get sharper at what they repeat.
MCP server
Connect Claude Code or any MCP host to your learnings and claims.
Claude Code
Install the hook that recalls learnings before your assistant acts.
Framework integrations
Add learnings to an agent you already run on LangGraph.
Learning loop over REST
Wire the learning loop into an agent you already run, in any language.
API reference
Claims search
Search the claims with a question in plain words and get back the ones nearest it, with their sources.
Claims API
List and read claims, settle conflicts, correct claims, fix who is who, and erase an entity.
Learnings API
Store, search, retrieve, and reinforce learnings.
Distill documents
Turn documents, diffs and post-mortems into learnings and claims.
Agent value API
Set what counts as value, record outcomes on an agent's ledger, and read its lifetime value.
Usage API
Inspect tenant and per-agent usage, costs, learning activity, and run history.
Platform
Access control (FGA and RBAC)
Scopes, organization roles, spaces, API key permissions, and the full permissions matrix.
Sharing and visibility
When one agent's learnings can reach another, what is stripped on the way, and why claims never travel.
Security and data isolation
Tenant isolation, scoped API keys, and access controls.
Who this is for
- Technical buyers evaluating whether Hyperstruck is more than a prompt wrapper
- Engineers looking for clean API entry points into learning and memory
- Teams that want to integrate Hyperstruck into existing agent frameworks or IDE workflows