Skip to content

introduction.mdx

Latest commit

 

History

History
118 lines (99 loc) · 7.46 KB

introduction.mdx

File metadata and controls

118 lines (99 loc) · 7.46 KB
Raw
title Introduction
description The trust plane for the agentic internet — transparent alignment verification, behavioral drift detection, and accountability protocols
sidebarTitle Introduction

The trust plane for the agentic internet

Mnemom makes AI agent behavior transparent and verifiable. When an agent makes a decision, Mnemom records what it decided, what alternatives it considered, what values it applied, and whether its reasoning was compromised. The result is a verifiable audit trail that any stakeholder can inspect.

Three layers power this:

  • AAP (Agent Alignment Protocol) -- Post-hoc verification. Alignment Cards declare an agent's values and boundaries. AP-Traces record every decision. Verification checks whether behavior is consistent with declared intent.
  • AIP (Agent Integrity Protocol) -- Per-turn integrity checking. Analyzes LLM thinking blocks during execution to detect prompt injection, value drift, and manipulation; in enforce it gates the response before the agent acts, while observe and nudge record the verdict post-hoc. AIP reads reasoning, not inbound content: the content itself is screened by Safe House at the gateway's front door, inside the request that carries it — including tool results returning from a tool call.
  • CLPI (Card Lifecycle & Policy Intelligence) -- Governance-as-code that enforces what tools agents may use, distinguishes configuration errors from behavioral failures, recovers trust after false violations, and anchors reputation on-chain.
**What Mnemom guarantees:** Complete audit trails of every agent decision (AAP), per-turn attestation of reasoning integrity at the thinking level (AIP), and active policy enforcement that blocks prohibited actions at the gateway in enforce mode (CLPI). Together, these provide verifiable accountability from declared intent through execution.

What no external system can guarantee: Enforcement at the sub-thinking level — inside the model's weights. If a model produces compromised reasoning that does not surface in its thinking blocks, no external observer can detect it. Mnemom verifies everything the model externalizes; it cannot inspect what the model does not reveal. See the limitations docs for details.

Get started

These quickstarts are available in [Spanish (Español)](/es/quickstart/overview) and [French (Français)](/fr/quickstart/overview) — six pages per language have been translated. **Recommended.** Wrap any LLM API with zero code changes. Full AAP and AIP compliance in minutes. Integrate AAP and AIP directly into your application for maximum control over tracing and verification. Run the Mnemom gateway, the API, and the database on your own infrastructure. Full data residency control.

Key concepts

A machine-readable declaration of an agent's values, autonomy boundaries, escalation triggers, and audit commitments. The "constitution" every trace is verified against. Structured records of agent decisions -- what action was taken, what alternatives were considered, what values were applied, and whether escalation was required. Per-turn AIP analysis of LLM thinking blocks. Each checkpoint produces a verdict: `clear`, `review_needed`, or `boundary_violation`. Statistical monitoring of agent behavior over time. Detects when an agent's actions gradually diverge from its declared alignment. Card Lifecycle & Policy Intelligence. Governance-as-code with policy enforcement, violation reclassification, trust recovery, risk intelligence, and on-chain reputation anchoring. A credit score for AI agents. Five-component composite metric computed from integrity checkpoints, drift stability, compliance, trace completeness, and fleet coherence. Cryptographically provable and anchorable on-chain. Teams are first-class meta-agents with persistent identity, their own alignment cards, and accumulated reputation — independent of any individual member. Four-layer cryptographic attestation: Ed25519 signatures, hash chains, Merkle proofs, and optional SP1 zero-knowledge proofs. Any party can independently verify a verdict without trusting Mnemom. Kill-switch for rogue agents. Pause, kill, and resume agents in real-time with auto-containment triggers, RBAC controls, and full audit trails. Pairwise compatibility checking between agents. Before two agents collaborate, coherence checks verify their values are not in conflict.

SDK packages

Install the protocol SDKs directly if you need fine-grained control:

```bash npm npm install @mnemom/agent-alignment-protocol npm install @mnemom/agent-integrity-protocol ``` 
pip install agent-alignment-protocol
pip install agent-integrity-proto
Package Language Registry Description
@mnemom/agent-alignment-protocol TypeScript npm AAP verification, coherence, drift detection
@mnemom/agent-integrity-protocol TypeScript npm AIP integrity checking, signal building
agent-alignment-protocol Python PyPI AAP verification, coherence, drift detection
agent-integrity-proto Python PyPI AIP integrity checking, signal building
**Canonical Python package names:** PyPI: `agent-integrity-proto` (AIP) and `agent-alignment-protocol` (AAP); npm: `@mnemom/agent-integrity-protocol` (AIP) and `@mnemom/agent-alignment-protocol` (AAP). These are the only published package names — do not install packages under any other name.

Protocols

How AAP and AIP work together as complementary verification layers. Full Agent Alignment Protocol specification for implementers. Full Agent Integrity Protocol specification for implementers. Threat models, attack surfaces, and known limitations.