Get StartedGuidesSoonAPI ReferenceSoonRelease NotesSoon
Core Concepts

How Restrct works

Restrct sits beneath your agent runtime as a single infrastructure layer - memory, guardrails, and harness engineering, all sharing one API key.

Memory Engine

Every write to restrct.memory.add() is scored, tiered, and indexed automatically. Hot memories stay in fast recall, cold ones age out into long-term storage without you managing eviction manually.

Governance Guardrails

Guardrails are policy rules evaluated before an agent action executes - scoped by workspace, role, or agent identity - so a misbehaving agent fails closed instead of silently doing the wrong thing.

Agent Harness

The harness wraps tool calls, retries, and context assembly so your agent code stays focused on decisions rather than plumbing. It works with Claude, Cursor, Codex, and any framework that speaks the Restrct SDK.