mirra-sdk

One wrapper call gives any agent a persistent recognized identity, tamper-evident portable memory of its relationships, behavior that adapts per relationship, and execution that only happens when verified.

import mirra

wrapped = mirra.wrap(my_agent, principal="team-key-1", profile="dev_balanced")

wrapped.identity                     # recognized across sessions, not met fresh
wrapped.remember("alice", "…")       # signed scroll — verify-on-read, fail-closed
wrapped.recall("alice")              # only cryptographically verified memories
wrapped.interact("alice", "hi")      # context built from alice's own history
wrapped.execute("shell.exec", "…")   # deterministic, signed decision record

This is Edge 1 (the SDK) of the MIRRA platform's one-core/three-edges architecture. The SDK speaks only the frozen v1 core contract (mirra-core-contract); the enforcement engine (MVAR via ClawZero) and the signed memory store (ClawSeal) plug in beneath it, and optional capability providers can be injected at runtime for identity enrichment and epistemic verification.

The four pillars

Profiles: build against dev_balanced (safe actions allow, hostile ones block, risky ones step up). The default is prod_locked — the fail-closed production posture where every action is refused unless the caller supplies risk context or step-up approval. If you omit profile and everything blocks, that is the lockdown working as designed.

Fail-closed by design

• No enforcement engine → every action is blocked, never silently allowed.
• No memory backend → wrap() raises, never runs with unsigned memory.
• Damaged identity keystore → raises, never silently mints a new identity.
• Tampered scroll or forged decision record → fails verification.

Demo

Installed (pip):

mirra-demo                   # live 5-check report; run twice — the second run proves recognition

From a checkout:

python3 demo/demo_sdk.py     # narrative version of the same proof

See DEMO.md for the five-minute walkthrough.

Tests

python3 -m pytest tests/ -q

The suite is organized by acceptance criterion: test_recognition.py, test_signed_memory.py, test_differentiated_behavior.py, test_permissioned_execution.py, test_boundary.py (public/private leak gate).