Compound Memory¶

Client: Vishal Gulati, Meridian (a Recode Ventures spinout).

Most artificial intelligence (AI) effort in drug discovery goes into designing new molecules. Meridian takes a different bet: Restless, an intelligence system that reads trial registries, scientific literature and patent databases, and produces structured dossiers on clinical programmes worth licensing. Each dossier, however, exists in isolation: what Restless establishes about a target, a company or a mechanism cannot compound across runs. Your task is to build Compound Memory, a knowledge store that ingests Restless's dossiers, recognises when different dossiers refer to the same drug, company or target under different names, and links them into a single graph. Every fact must carry its source and date, so contradictions between runs stay visible. Compound Memory serves two clients: a human analyst asking a defined set of structured questions, and an AI agent retrieving a compact context bundle before a new run. Success is measured by linking accuracy on a hand-checked sample and by analyst questions answered correctly. If it works, a small team can track far more programmes in diseases where no effective treatment exists. Restless's output schema, the interface specification Compound Memory must implement, a glossary and sample dossiers will be provided at the start.