Precedent — Every experiment should make the next one smarter. What if the next breakthrough in research didn't have to start from scratch? AI agents can run experiments and discover what works, but those discoveries often remain isolated. When another agent tackles a similar problem, it may repeat failed approaches, overlook useful results, or apply findings under the wrong conditions. Precedent is a shared, governed research memory that allows AI agents to build on each other's experimental experience. Rather than simply remembering conversations, Precedent stores structured experimental findings: what was attempted, what happened, under which conditions, and which model checkpoints were produced. Other agents can search this evidence, evaluate whether it applies to their task, and use it to make better decisions. In our live demonstration, Codex has one four-epoch run to train a ResNet-50 model to reach at least 80% accuracy on unseen handwritten digits. It searches previous experiments through Precedent, identifies a successful checkpoint, avoids a previously unsuccessful approach, and excludes a result from a different image resolution. A human reviews and approves the proposed plan before execution. Precedent retrieves the saved model weights, runs the experiment, and publishes the new finding back to MongoDB Atlas for the next agent. In one measured comparison, the scratch baseline achieved 29.2% accuracy, while the checkpoint-assisted run achieved 90.4% on the same 240-image holdout. These figures were measured once. Precedent turns isolated experiments into shared, reusable evidence—so one agent's discoveries can become another agent's starting point.
Team
