BrainBob - Shared Memory

BrainBob is a shared brain for agent teams: keep knowledge current, control who can use it, and trace every change. When one agent remembers an outdated deadline, the next agent can repeat it with confidence. BrainBob checks incoming facts against existing memories, applies scope and authority rules, and supplies fresh agents with relevant, current context. During the hackathon, we built a live agent workspace with personal chat and a 3D memory office. Users can inspect retrieved files and citations, follow source history, correct outdated records, and retire memories while preserving their history. A dedicated NVIDIA Nemotron Nano model generates summaries of each person’s accessible memory, alongside the original records. We also built the Bob organizational-memory benchmark. Across six synthetic worlds, 41 checkpoints, and three replicates, maintained BrainBob memory scored 71.9% on our composite all-checks-pass metric versus 25.0% without memory. These are corrected v0.3 scores from rescoring saved answers; the advantage over append-only memory remains inconclusive. Separately, maintenance correctly updated 12/13 superseded facts versus 0/13 for append-only memory, without wrongly retiring any of 32 valid memories. Our NemoClaw/OpenClaw integration includes MCP tools for reading memories, inspecting sources, and proposing new knowledge. The adapter is implemented and unit-tested; execution inside a NemoClaw/OpenShell sandbox remains unverified.