Context
infrastructure
for AI agents
Restrct compiles official documents and accumulated team knowledge into one deterministic, source-labeled context package, so every agent answer traces back to where it came from.
Restrct compiles official documents and accumulated team knowledge into one deterministic, source-labeled context package, so every agent answer traces back to where it came from.
Powering AI infra for
AI agents need context, but it often comes from scattered, disconnected sources. Restrct compiles team knowledge and official documents into a single deterministic, source-labeled context package that agents can act on safely.
Ingests official documents (policies, contracts, compliance obligations) and extracts structured operational facts. Permissions, business rules, workflows, risks, and compliance obligations become the authoritative source of truth that always outranks informal memory when the two disagree.
A persistent, multi-tenant memory layer for AI assistants. It captures facts, events, and relationships from anything you tell it, then makes them instantly recallable - so your assistant remembers who you are, what you've decided, and how things have changed over time. It skips duplicate noise, tracks how facts change, and links related people, projects, and decisions
Searches Knowledge and Memory in parallel, merges them into one ranked, source-labeled result, and flags anywhere the two disagree. Every piece of output carries its origin - policy or memory, binding or background - end-to-end. And because no LLM sits in the retrieval path, the same query against the same data returns the same result every time, so you can trace and reproduce any answer.
Export to LangGraph, CrewAI, AutoGen, OpenAI Agents SDK, or any REST API. Restrct sits underneath your agent stack as the context layer, your framework of choice on top, a deterministic and auditable context layer underneath.
import { Restrct } from '@Restrct/sdk'
// Ingest your compliance docs
const blueprint = await Restrct.Blueprint.ingest({
documents: ['compliance.pdf', 'sop.docx'],
framework: 'langgraph',
})
// Guardrails auto-generated from your docs
console.log(blueprint.guardrails.length) // e.g. N rules
console.log(blueprint.tests.length) // e.g. N cases
// Deploy an auditable, source-labeled agent
const agent = await blueprint.deploy()Two sources go in. One deterministic, source-labeled context package comes out.
Feed in official documents (policies, contracts, compliance obligations) alongside team knowledge like past decisions and institutional memory.
Structured extraction turns your documents into operational facts: permissions, business rules, compliance obligations, risks, and workflows. This becomes the authoritative source of truth.
A persistent, per-user memory layer. Tracks how things change over time reconstructs timelines from dated events, and builds a relationship graph of people, projects, and technologies - blended into one result so the right memory surfaces no matter how you ask.
Knowledge and Memory are searched in parallel. Results are merged into one ranked, source-labeled package. Anywhere the two disagree, the conflict is flagged explicitly.
The final output is compiled as a clear split: binding instructions versus background context. Points requiring human approval are flagged automatically before the agent acts.
Knowledge, Context, and Harness run with no LLM: same query, same data, always the same result. Every piece of output keeps its source label end-to-end, auditable at every step.
An AI agent is only as reliable as the context beneath it. We built Restrct so every answer is source-labeled and reproducible - auditable by design, not as an afterthought.
Every piece of context Restrct produces carries its source end-to-end: policy document or team memory, binding instruction or background context. When the two disagree, official documents always win. Points requiring human approval are flagged before the agent acts, not left for the agent to decide. And because there's no LLM in Knowledge, Context, or Harness, the same query against the same data always produces the same result, which means your compliance team can actually audit it.
Start on an auditable, source-labeled substrate from day one - no infrastructure to build.