# I Built a Personal AI Agent on a Raspberry Pi — Jeremy Adams, Neo4j

## Executive summary

Jeremy Adams details the construction of a wearable, personal AI agent, NanoClaw, running on a Raspberry Pi 4B. This system demonstrates an offline-first approach to agent memory, utilizing a local Neo4j graph database and the POLE schema (Person, Object, Location, Event) to capture and enrich real-world interactions (e.g., conference booth notes). The architecture connects the physical device to cloud LLMs (Claude) and a centralized Neo4j graph database via WhatsApp messaging, allowing for persistent, context-aware memory capture even without constant internet connectivity.

## Key takeaways

- Wearable, Hackable Agents: The speaker advocates for small, open, and hackable edge devices (like the Raspberry Pi) over feature-rich, proprietary laptop agents, prioritizing understanding and customization.
- Graph Memory with POLE Schema: Memory is built using the POLE schema (Person, Object, Location, Event) to structure captured data, allowing the agent to build a connected graph of experiences.
- Offline-First Architecture: The system uses a local Neo4j instance on the Raspberry Pi for offline data capture, which is later synced and enriched in the cloud, ensuring continuity regardless of network quality.
- System Integration Pipeline: The agent processes input via WhatsApp messaging, runs local inference/capture on the Raspberry Pi, and sends queries/data to the cloud Neo4j instance via an MCP server connection.

## Technical details

- Hardware & Edge Computing: The agent runs on a Raspberry Pi 4B, demonstrating capability for complex tasks on low-power, hackable hardware. The setup includes a USB microphone and a custom-built physical button trigger.
- Software Architecture: The system utilizes Docker containers to isolate agent processes and runs Neo4j on the device. The initial messaging pipeline connects WhatsApp (cloud) $\rightarrow$ NanoClaw (Pi) $\rightarrow$ Claude (cloud LLM).
- Graph Database Implementation: Neo4j is used for both the core database and memory storage. The POLE schema is implemented to model memories, allowing the agent to connect disparate data points (e.g., a person, a location, and an event).
- Data Capture & Enrichment: The agent captures raw notes (e.g., conference booth details) locally. These raw notes are later uploaded and enriched in the cloud, connecting exhibitors to broader 'theme nodes' (e.g., Evaluation, Observability).

## Practical implications

- The design demonstrates a robust pattern for building edge-computing AI agents that function reliably in low-connectivity environments.
- The use of graph databases (Neo4j) for memory management provides a structured, highly interconnected way to store and retrieve contextual knowledge from unstructured inputs (like notes or conversations).
- The modular, containerized approach (Docker) allows for easy modification and scaling of agent components on limited hardware.
- This architecture provides a blueprint for field data collection systems that require local processing before cloud synchronization and enrichment.

## Topics

Graph Databases, Edge AI, Agent Architecture, Neo4j, POLE Schema, Offline Computing, Docker, NanoClaw, Neo4j Agent Memory, GraphAcademy

Source: https://www.youtube.com/watch?v=oUZEt4EiPbk