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

Jeremy Adams, a senior developer advocate at Neo4j, runs his personal AI agent on an old Raspberry Pi that he wears around his neck. In this talk from the AI Engineer World's Fair 2026 in San Francisco, he assembles it on stage.

After a short intro and a hello from Neo4j (0:12), he lays out what he is building at 1:32. The live hardware setup starts at 1:57 and runs about three minutes, and the Pi boots up at 4:51. According to the description, the board is a Raspberry Pi 4B running NanoClaw, with agents isolated in Docker.

The reasoning comes next. At 6:06 he argues for small claws, and the description sums up his case: a small, hackable claw beats a feature-rich agent on your laptop. At 8:05 he explains why he chose NanoClaw. The architecture at 8:55 runs from WhatsApp to NanoClaw to Claude.

Memory takes up the middle of the talk. A short graphs primer at 9:40 leads into querying Neo4j over MCP at 10:10. At 10:50 he describes building graph memory on a plane, using the POLE schema (person, object, location, event), and he demos the resulting memory graph at 12:15.

Voice-to-text arrives at 13:29, added through a USB microphone and a hardware button. At 14:49 he records notes at the conference expo booths. Offline mode with a local Neo4j instance is at 15:34, and at 17:03 the booth notes are enriched into themes in the cloud. At 18:23 he upgrades the setup to Neo4j Agent Memory. The talk ends at 19:33 with a pointer to GraphAcademy, which the description lists as free courses.