@mnemom/agent Alignment Protocol

Agent Alignment Protocol (AAP)

A transparency protocol for autonomous agents.

AAP lets agents declare their alignment posture, produce auditable decision traces, and verify value coherence before coordinating with other agents. It extends existing protocols (A2A, MCP) with an alignment layer that makes agent behavior observable.

AAP is a transparency protocol, not a trust protocol. It makes agent behavior more observable, not more guaranteed.

Quick Start

# Install
pip install agent-alignment-protocol

# Generate an Alignment Card
aap init --values "principal_benefit,transparency,harm_prevention"
# ✓ Created alignment-card.json

# Instrument your agent

from aap import trace_decision

@trace_decision(card_path="alignment-card.json")
def recommend_product(user_preferences):
    # Your agent logic here
    # Decisions are automatically traced
    ...

# Verify behavior matches declaration
aap verify --card alignment-card.json --trace logs/trace.json
# ✓ Verified [similarity: 0.82]
# Checks: autonomy, escalation, values, forbidden, behavioral_similarity

Why AAP?

The agent protocol stack provides capability discovery (A2A), tool integration (MCP), and payment authorization (AP2). None address a fundamental question: Is this agent serving its principal's interests?

As agent capabilities become symmetric—equal access to information, equal reasoning power—alignment becomes the primary differentiator. AAP provides the infrastructure to make alignment claims verifiable.

Three Components

┌─────────────────┬─────────────────┬─────────────────┐
│ Alignment Card  │    AP-Trace     │ Value Coherence │
│                 │                 │    Handshake    │
├─────────────────┼─────────────────┼─────────────────┤
│ "What I claim   │ "What I         │ "Can we work    │
│  to be"         │  actually did"  │  together?"     │
└─────────────────┴─────────────────┴─────────────────┘
     Declaration        Audit          Coordination

Alignment Card

A structured declaration of an agent's alignment posture, in the unified / ADR-039 shape accepted by mnemom card validate and the Mnemom platform:

{
  "card_version": "unified/2026-04-26",
  "card_id": "ac-my-agent-001",
  "agent_id": "did:web:my-agent.example.com",
  "issued_at": "2026-04-26T00:00:00Z",
  "autonomy_mode": "enforce",
  "integrity_mode": "enforce",
  "principal": {
    "type": "human",
    "identifier": "did:web:user.example.com",
    "relationship": "delegated_authority"
  },
  "values": {
    "declared": ["principal_benefit", "transparency", "minimal_data"],
    "conflicts_with": ["deceptive_marketing", "hidden_fees"]
  },
  "autonomy": {
    "bounded_actions": ["search", "compare", "recommend"],
    "escalation_triggers": [
      {
        "condition": "purchase_value > 100",
        "action": "escalate",
        "reason": "Exceeds autonomous spending limit"
      }
    ],
    "forbidden_actions": ["share_credentials", "subscribe_to_services"]
  },
  "audit": {
    "trace_format": "ap-trace-v1",
    "retention_days": 90,
    "queryable": true,
    "query_endpoint": "https://my-agent.example.com/api/traces"
  }
}

Migration note (v2.0.0): earlier releases used the AAP 0.5.0 card shape (aap_version, autonomy_envelope, audit_commitment). The card has moved to the unified / ADR-039 shape above: autonomy_envelope → autonomy, audit_commitment → audit, aap_version → card_version, plus the top-level autonomy_mode / integrity_mode master switches and a required principal.identifier when principal.type != "unspecified". See the Alignment Card Schema.

AP-Trace

An audit log entry recording each decision:

{
  "trace_id": "tr-f47ac10b-58cc-4372",
  "card_id": "ac-f47ac10b-58cc-4372",
  "timestamp": "2026-01-31T12:30:00Z",
  "action": {
    "type": "recommend",
    "name": "product_recommendation",
    "category": "bounded"
  },
  "decision": {
    "alternatives_considered": [
      {"option_id": "A", "score": 0.85, "flags": []},
      {"option_id": "B", "score": 0.72, "flags": ["sponsored_content"]}
    ],
    "selected": "A",
    "selection_reasoning": "Highest score. Option B flagged as sponsored and deprioritized per principal_benefit value.",
    "values_applied": ["principal_benefit", "transparency"]
  },
  "escalation": {
    "evaluated": true,
    "required": false,
    "reason": "Recommendation only, no purchase action"
  }
}

Value Coherence Handshake

Pre-coordination compatibility check between agents:

from aap import check_coherence

result = check_coherence(my_card, their_card, task_context)

if result.compatible:
    # Proceed with coordination
    proceed_with_task()
else:
    # Handle conflict
    print(f"Value conflict: {result.conflicts}")
    # Escalate to principals or negotiate scope

What AAP Does Not Do

This matters. Read it.

1. AAP does NOT ensure alignment—it provides visibility. An agent can produce perfect traces while acting against its principal's interests.

2. Verified does NOT equal safe. A verified trace means consistency with declared alignment. It doesn't mean the alignment is good or the outcome was beneficial.

3. AP-Trace is sampled, not complete. Traces capture decision points, not every computation. Significant reasoning may occur between traces.

4. Value coherence is relative to declared values. The handshake checks if declared values are compatible. It doesn't verify agents hold these values or will act on them.

5. Tested on transformer-based agents. Other architectures may exhibit behaviors AAP doesn't capture.

For the complete limitations disclosure, see Section 10 of the Specification.

Installation

# Python
pip install agent-alignment-protocol

# TypeScript
npm install @mnemom/agent-alignment-protocol

Integration

With A2A Agents

AAP extends the A2A Agent Card with an alignment block:

{
  "name": "Shopping Assistant",
  "description": "Helps users find products",
  "url": "https://shopping.example.com",
  "alignment": {
    "$ref": "./alignment-card.json"
  }
}

See A2A Integration Guide.

With MCP Tools

AAP can trace tool invocations and verify they match declared alignment:

from aap import mcp_traced

@mcp_traced(card_path="alignment-card.json")
def my_tool(params):
    ...

See MCP Migration Guide.

Verification

The verification engine checks whether observed behavior matches declared alignment:

from aap import verify_trace, detect_drift

# Single trace verification
result = verify_trace(trace, card)
print(f"Verified: {result.verified}, Similarity: {result.similarity_score:.2f}")
if not result.verified:
    print(f"Violations: {result.violations}")
if result.warnings:
    print(f"Warnings: {result.warnings}")

# Drift detection over time
alerts = detect_drift(card, recent_traces)
for alert in alerts:
    print(f"Drift detected: {alert.analysis.drift_direction}")

Verification checks:

• Autonomy compliance (actions within declared bounds)
• Escalation compliance (required escalations were performed)
• Value consistency (applied values match declared values)
• Forbidden action compliance (no forbidden actions taken)
• Behavioral similarity (semantic alignment using SSM analysis)

Action Name Matching

Bounded and forbidden action lists support flexible matching (v0.1.7+):

{
  "bounded_actions": [
    "search",
    "exec: execute shell commands",
    "cron: manage scheduled patrol jobs"
  ]
}

• Exact match: "search" matches action name "search"
• Colon-prefix match: "exec: execute shell commands" matches action name "exec"
• Compound actions: action name "exec, read" matches when each component is in the bounded list

This allows descriptive entries in alignment cards while keeping verification flexible.

Similarity scoring: Each verification returns a similarity_score (0.0-1.0) measuring semantic similarity between the trace and declared alignment. If a trace passes structural checks but has similarity_score < 0.50, a low_behavioral_similarity warning is generated.

Try It

Interactive Playground — Verify traces in your browser with SSM visualization.

• Paste your Alignment Card and AP-Trace
• See verification results with similarity scoring
• Visualize behavioral patterns with SSM heatmaps
• Adjust thresholds in real-time

No server required — runs entirely client-side via WebAssembly.

Documentation

Examples

Status

Current Version: (Python) / (TypeScript)

API Reference

# Core API
from aap import (
    verify_trace,      # Verify single trace against card → VerificationResult
    check_coherence,   # Check value compatibility between agents → CoherenceResult
    detect_drift,      # Detect behavioral drift over time → list[DriftAlert]
    trace_decision,    # Decorator for automatic AP-Trace generation
    mcp_traced,        # Decorator for MCP tool tracing
)

# Models
from aap import (
    AlignmentCard,
    APTrace,
    VerificationResult,  # .verified, .similarity_score, .violations, .warnings
    CoherenceResult,     # .compatible, .score, .value_alignment
    DriftAlert,          # .analysis.similarity_score, .analysis.drift_direction
)

# CLI
# aap init [--values VALUES] [--output FILE]
# aap verify --card CARD --trace TRACE        → Shows [similarity: X.XX]
# aap check-coherence --my-card MINE --their-card THEIRS
# aap drift --card CARD --traces TRACES_DIR   → Uses SSM analysis

Standards & Compliance

AAP aligns with and supports compliance for the following international standards and regulatory frameworks:

Contributing

We welcome contributions. See CONTRIBUTING.md for guidelines.

Key areas where we need help:

• SDK implementations in other languages
• Integration examples with popular agent frameworks
• Test vectors for edge cases
• Documentation improvements

License

Apache 2.0. See LICENSE for details.

Agent Alignment Protocol — Making agent alignment observable.