Wearable AI Agent Built on Raspberry Pi

The project centers around a wearable AI agent powered by a Raspberry Pi, and Adams engaged the audience to gauge their familiarity with edge devices, graph databases, and agent memory concepts.

A Raspberry Pi 4B, USB microphone, and video capture device formed the components for the live assembly, which Adams described as his most hands-on demonstration to date.

The hardware components included a Raspberry Pi 4B, a USB microphone for audio input, and a video capture device to manage visual data.

He then initiated the boot sequence on stage to confirm the agent's immediate functionality after assembly.

It operates on a 64-bit ARM operating system, a significant upgrade from Adams' previous 32-bit setup, and is supported by a portable battery for power, along with a Bluetooth keyboard for user interaction.

Jeremy Adams explained that his personal AI agent system runs NanoClaw, Docker, and a local Neo4j database directly on the Raspberry Pi 4B.

He prioritized creating an open, hackable, and low-cost solution over a feature-rich, black-box system, aiming to understand the underlying mechanisms rather than relying on opaque pre-installed software.

The setup represents an evolution from a bulky desktop configuration to a compact, portable device, reflecting a deliberate shift towards a more minimalist and integrated design.

Adams selected NanoClaw for his project due to its compact nature, comprising only 15 source files, which suited his preference for small and manageable codebases.

The architecture employs Docker containers for isolating agent processes and utilizes WhatsApp for messaging, allowing the Raspberry Pi to retrieve messages from the cloud; all large language model inference is conducted in the cloud via Claude, with the local hardware managing messaging and logic.

He illustrated this with an example where a 'Tom Hanks' person node connects to a 'Forrest Gump' movie node, emphasizing how an MCP server links the agent to the Neo4j instance.

The agent leverages this graph structure to retrieve specific information from natural language queries received via WhatsApp.

The agent's operational process involves fetching relevant data directly from the Neo4j database to inform its responses.

This information is then routed through a cloud agent SDK for processing and response formulation, with the final output delivered back to the user via the WhatsApp channel.

This schema was inspired by European policing methods, which have long used graphs for tracking and organizing data effectively.

A key feature of this memory system is its ability to persist data across system reboots, ensuring continuous recall and learning for the AI agent.

Examples included connections between the Snowflake office and specific AI meetup events, with individuals like Rebecca and Jeremy logged as distinct entities.

This system effectively functions as a digital record, capturing professional activities and their contextual links in a structured format.

This functionality is triggered by a physical button on the device, providing a tactile interface for interaction.

LED lights are incorporated into the system to offer clear visual feedback to the user during active recording sessions.

Adams created a simple data model of exhibitors by recording booth messaging at an expo, including basic data cleaning to organize the information within the database.

He implemented an offline mode using a local Neo4j instance to ensure continuous data collection despite potentially unreliable conference Wi-Fi, which allowed him to gather information without an active internet connection.

Voice-to-text functionality was utilized to parse booth numbers and insert them into the local database using regular expressions, demonstrating an efficient method for data entry during the event.

Adams showcased a lightweight tool he developed, the 'Cipher Shell browser,' which allows querying Neo4j databases without requiring a complex graphical user interface.

He demonstrated its capability by matching and retrieving exhibitor data directly from the device worn around his neck, explaining that raw notes from the event were captured and later uploaded to a cloud-based Neo4j instance for further processing and enrichment.

Enriched cloud data revealed significant connections between various exhibitors and emerging industry themes, such as 'Evaluation' and 'Observability,' linked to companies like Buildkite and LangChain.

This process allowed Adams to generate novel insights by mapping how disparate companies relate to common technical trends, highlighting underlying thematic relationships that were not immediately obvious from raw data.

This service distills raw chat history into structured memories, categorizing information about people, locations, and concepts.

These distilled memories are then made accessible via an MCP server, providing deeper context for subsequent agent interactions.