NeuroRecursion

Brain-computer interfaces lose accuracy every day as neural signals drift, so today’s systems make users recalibrate. Neuralink users spend about 55 minutes a week labelling data so part of the decoder can be retrained. NeuroRecursion freezes the decoder completely and puts a recursive agent team in front of it. Each day, a Nemotron planner on Crusoe turns the goal “keep the cursor usable without recalibration” into checks it can run itself. Tool agents (channel health, signal alignment, intent check) propose small, reversible fixes, and a label-free verifier accepts or rejects each one. The agents build a growing memory of the user. Between versions, a Nemotron reviewer reads the audit log and shared memory and rewrites the team’s playbook. A change is kept only if it scores better on drift scenarios the team has never seen. Results: On Neuralink’s public raw recordings, spike detection cuts data about 1,600×, enough for the implant’s radio. On public monkey motor-cortex data (MC_Maze) with simulated drift, the agent system matched the existing approach (99% frozen model plus daily labelled retraining) for the first three days. It pulled ahead on days 4–7 as its memory grew (+10 points, 3 of 5 unseen scenarios) with no labels at all. That’s a trend, not yet statistically significant. In the Version N+1 loop, the Nemotron reviewer found one playbook edit that improved held-out decoding (R² 0.043 → 0.059), and the gate reverted every edit that didn’t help. All code, evaluations and 3D visualisations were built during the hackathon. This is a research prototype, not a medical device.