I built a unified memory and MCP gateway for AI agents that syncs in ~12ms. Here is how it works.

Building autonomous workflows often led to a phenomenon known as "framework fatigue." Switching between different development environments, such as coding tools, command-line interfaces, or browser-based assistants, required developers to repeatedly configure tools, authenticate keys, and re-index files. This process was not only tedious but also fragile and lacked scalability.

To address these issues, the author developed Memorify, a unified memory and Model Context Protocol (MCP) gateway designed to streamline AI agent workflows.

Memorify aimed to serve as a single entry point for all AI agents, abstracting vector storage, memory management, and OAuth connections behind a unified interface. The key features of Memorify included a frictionless developer onboarding process, persistent MCP hot-plugging, and a verb-based semantic protocol that minimized schema complexities.

Upon registering an agent, Memorify issued a secure Bearer token for authentication. This token was then used in a JSON-RPC POST request to retrieve a list of available tools registered under the agent's profile. By simply calling the 'whoami' tool, the agent could verify its identity, workspace details, and authorization scopes, thereby establishing a secure connection.

A standout feature of Memorify was its ability to support Persistent MCP Hot-Plugging. This meant that developers could add, update, or remove MCP servers without disrupting the agents' flow. Changes made through the Memorify dashboard were instantly reflected across all connected agents, eliminating the need for reboots or reloads.

In terms of architecture, Memorify adopted a verb-based semantic protocol, which contrasted with traditional schema-heavy database interactions. Agents communicated using natural language intents, focusing on actions like remembering, recalling, and updating information. Memorify supported four main capabilities: /memory (for storing and retrieving facts), /documents (for ingesting and searching content), /skills (for executing modular tools), and /mcp (for routing MCP requests).

The execution stack of Memorify was designed for low latency and high availability. Key components included Deno Deploy, which provided edge routing with minimized cold-start penalties; Neon Postgres, which handled serverless relational data and native vector storage, delivering instant database scaling and high-performance vector searches; and ElectricSQL, which managed active-active synchronization of database replicas for near real-time updates.

This architecture enabled instant visibility of actions taken in one agent to others, ensuring a cohesive workflow experience across multiple AI agents.