Hermes Framework in Practice: Building a Self-Evolving AI Shopping Assistant

In the wave of AI agents evolving from "task tools" to "digital employees," the Hermes framework stands out for its persistent memory and automatic skill generation capabilities. Based on a live demo by Nihal Kumar Khatri, an AI engineer at Infosys, presented at a DataCamp technical workshop, this article breaks down how to use Hermes to build an AI shopping assistant with memory, autonomous learning, and cross-platform interaction.

Core Differences Between Conventional Agents and Hermes Agents

To understand the value of Hermes, we first need to recognize its fundamental differences from traditional agents. Conventional AI agents are stateless — you chat with them, but once the conversation ends, all memory disappears. If you tell it "I wear a size 10" or "I prefer a certain material," it will have no recollection of that the next time you use it.

Hermes, by contrast, implements persistent memory. In the demo, when a user shares their preferred size and budget, that information is permanently stored in the system. More critically, this memory can be shared across conversation threads — a user can close the current session, start a brand-new thread, and ask "What's my usual size and budget?" Hermes will still answer accurately, because memory is stored in a dedicated folder under the user's directory rather than being tied to a single session.

Another fundamental difference lies in the skill mechanism. Conventional agents have skills that are hand-coded by developers for specific scenarios; Hermes is a continuously self-evolving agent — its skills are built at runtime. After making a mistake or completing a complex task, it learns from the experience and generates reusable skills that are stored in a database for use on similar future tasks.

System Architecture: An Integrated Multi-Tool, Multi-Platform Design

In this demo setup, Hermes serves as the central brain, and the overall architecture is highly extensible.

Flexible Deployment Options

Hermes can run via a local terminal, be deployed on an Oracle Cloud VM for 24/7 hosting, or be integrated with mainstream chat apps like Telegram, Discord, WhatsApp, and Slack. This means users no longer need to switch between multiple shopping sites — they can search, compare prices, add items to cart, and even check out directly within a familiar chat interface.

A Completely Free Model Stack

To keep costs down, the demo uses an entirely free model combination: an open-source model from Nous Research by default, along with a fallback model via OpenRouter — if the primary model hits a rate limit, the system automatically switches to a backup (such as Tencent's Hunyuan model), ensuring service continuity.

Three Core Tool Capabilities

The solution integrates two FastMCP-based servers and a web search capability:

• Local Store MCP: The product catalog is written directly in code for fast retrieval of local product data
• Shopify MCP: Inventory items loaded into a real online store, supporting full cart operations
• Real-time Web Search: Can directly search e-commerce sites like Amazon and Walmart and scrape product information

Live Demo: The Complete Flow from Query to Checkout

Intelligent Routing and Multi-Tool Invocation

When a user asks for "a pair of Nike running shoes in size 10 under $100," Hermes intelligently routes the request to the appropriate tools. In the demo, it simultaneously called Store Search Products (local catalog) and Shopify Search Products (real inventory), finding 2 matching pairs from the local store and 3 from the Shopify store. This behavioral logic is clearly defined in the configuration file, ensuring multi-tool invocations behave exactly as expected.

Storing and Retrieving Memories

After adding items to the cart, the user provided their preferred size and budget, which Hermes saved to persistent memory. There is a key configuration flag here — the memory storage flag. If disabled (set to False), the system will not retain memory, returning control to the user to accommodate varying privacy preferences.

Automatic Skill Generation

The most impressive part of the demo was automatic skill generation. When the user first asked to compare two shoes (Nike Zoom Vega vs. an Adidas model), Hermes completed a multi-step reasoning analysis and saved the "product rating comparison" process as a reusable skill, stored as a Markdown file.

Interestingly, skill generation can be fully automated. The system's default trigger logic is: automatically generate and store a skill whenever more than 5 tools are invoked in a single query. However, to introduce a Human-in-the-Loop mechanism, the solution provides a "write approval" flag — set to True for AI to dynamically generate skills at runtime, or False to skip saving and require human review and approval. This design is especially valuable in production environments, effectively balancing automation efficiency with quality control.

Cross-Platform Deployment and Autonomous Scheduling

A Real Shopping Experience on Telegram

Once the solution is deployed to an Oracle VM, users can complete the entire shopping flow through Telegram: search for products, add to cart, and generate a checkout link. Clicking the link redirects to the Shopify checkout page, where address information is pre-configured and the order can be placed by entering payment details.

The entire payment step is handled by a UCP (Universal Commerce Protocol) endpoint — a unified standard co-developed by Google with Shopify and Walmart. Shopify is migrating from traditional MCP to UCP, representing the future direction of e-commerce protocols.

Cron Jobs: Letting the AI Agent Run Autonomously

Hermes supports automated scheduling via Cron Jobs. The "PriceWatch" task in the demo periodically checks price changes for tracked items — if a pair of shoes drops from $80, the system triggers an alert within minutes, notifying the user via Telegram or WhatsApp. This truly enables an AI shopping assistant that "runs autonomously once configured."

Key Discussions on Payments and Extensibility

Why Payments Remain a Challenge

When an audience member asked whether Hermes could directly handle payments, the presenter candidly acknowledged that this is intentionally avoided for security reasons. Payment is a highly deterministic process that cannot tolerate AI "hallucinating" details, so there is currently no out-of-the-box AI payment solution — payment processing still relies on the shopping sites themselves.

Does It Work with Any E-Commerce Platform?

On the question of whether non-Shopify store products could be purchased, the presenter noted that Amazon and Walmart each have their own MCP servers, and the industry trend is rapidly converging toward the unified UCP standard. In theory, as long as an e-commerce site has implemented an MCP server, Hermes has the extensibility to interact with it; once UCP matures, a unified commercial protocol for shopping assistants will emerge.

More Production Use Cases

Hermes isn't limited to personal shopping — it can be applied to a variety of production scenarios:

• Autonomous customer support: The agent learns and saves solution skills, progressively improving service quality
• Code review and bug fixing: Understands fix patterns and automatically handles recurring issues
• Data extraction and scheduled reporting: Automates data collection and generates analytical reports
• Incident response: Monitors anomalies in real time and triggers remediation workflows

The core value proposition: the more it's used, the smarter it gets — achieving true self-evolution through a built-in learning loop.

Conclusion

The Hermes framework represents an important direction in AI agent development: from passive execution to active memory, from hand-coded skills to self-evolution. While engineering challenges remain — such as payment security and maintaining login states (e.g., periodic manual re-login required after deployment) — the persistent memory, cross-platform interaction, and automated skill capabilities it demonstrates are clearly sketching out the early form of a "digital employee."

As the presenter noted, completing all your shopping through Telegram may still be a niche experience, but it very likely represents where AI interaction is headed.