Show HN: Trust Protocols for Anthropic/OpenAI/Gemini
Mnemom.ai introduces open-source protocols, AAP and AIP, designed to bring integrity and alignment to complex AI agent systems. Facing challenges in managing autonomous agents, the author developed these tools to define what agents can do, what they are thinking, and to proactively flag deviations. This solution addresses a critical gap in AI agent management, offering a technical approach to ensure agents adhere to intended behaviors and values.
The Lowdown
Mnemom.ai introduces a critical missing piece in the burgeoning world of AI agents: trust protocols. As multi-agent systems become more complex and autonomous, ensuring these agents adhere to intended behaviors and values is paramount. The author, facing this challenge in their own work, developed two open-source protocols—the Agent Alignment Protocol (AAP) and the Agent Integrity Protocol (AIP)—to provide behavioral contracts and runtime integrity monitoring, moving beyond reactive logging to proactive integrity enforcement.
- Agent Alignment Protocol (AAP): This protocol defines what an agent can do and has done through an "Alignment Card." This structured declaration specifies permitted/forbidden actions, escalation triggers, and core values that guide the agent's decisions.
- Agent Integrity Protocol (AIP): AIP monitors what an agent is thinking about doing and is allowed to do. It uses "Integrity Checkpoints" to compare an agent's real-time reasoning trace against its Alignment Card, proactively flagging boundary violations before actions are taken.
- Proactive Enforcement: Unlike traditional observability tools that report what happened, AIP intervenes proactively. For instance, if a customer support agent considers accessing forbidden payment data, AIP detects this intent and prevents the action, nudging the agent back in line before any harm is done.
- Multi-Agent Coordination: For collaborative multi-agent systems, AAP and AIP facilitate "Value Coherence." Agents exchange Alignment Cards to verify compatibility, surfacing potential conflicts (e.g., an agent prioritizing "move fast" clashing with one valuing "rollback safety") before coordination begins.
- Observability by Design: Both protocols are built for observability, emitting OpenTelemetry traces for every integrity check and verification. This allows seamless integration with popular observability platforms like Grafana, Datadog, and Langfuse, offering real-time insights into agent behavior.
- Transparency, Not Blind Trust: The creators explicitly clarify that these are "transparency protocols, not trust protocols." They aim to make agent behavior and reasoning observable but do not guarantee agents will always behave as declared, catch sophisticated deception, or replace human judgment.
- Open-Source & Accessible: The protocols are Apache-licensed, compatible with agents from Anthropic, OpenAI, and Gemini, and available as SDKs on npm and PyPI. A free gateway proxy (smoltbot) is also provided, enabling integrity checking with zero code changes for existing agents.
In essence, Mnemom offers a foundational layer for managing the burgeoning complexity of AI agent ecosystems. By providing standardized methods to define, monitor, and enforce agent behavior and values, these protocols aim to build a more transparent and manageable future for AI, fostering better control and understanding of autonomous systems.