News
Harvey said context graphs improved Legal Agent Bench rubric pass rates by up to 30% and cut task cost by about 50% in company tests.
Harvey said it is using graph-based context to improve legal-agent accuracy, arguing that the key differentiator for AI in legal work is not only model reasoning but the matter-specific information available before a task begins. In a company blog post published Oct. 19, 2026, Harvey described the approach as Harvey Graph and said it is designed to organize firm data, legal knowledge and their relationships so agents can start with relevant context, documents and precedents already in view.
Evidence and context
According to Harvey’s Oct. 19, 2026 post, the system combines a knowledge graph for U.S. case law with a context graph for per-Space matter data, documents and integrations. The company said its U.S. Case Law knowledge source maps citations and relationships across millions of opinions from federal district and appellate courts and all 50 states, while the context graph is intended to track how matters, documents, people and related work connect within a firm’s governed environment.
Harvey also said the graph is meant to reduce the need for repeated re-briefing. The company described scenarios in which an agent can identify the controlling version of an agreement, trace a diligence issue to the relevant clause and counterparties, and follow citation treatment rather than relying on flat retrieval that ranks text chunks by similarity alone. Harvey said the graph layer operates automatically for users, with no separate management required.
Research results and limitations
In experiments on Legal Agent Bench, Harvey said an agent using the context graph improved rubric pass rate by up to 30% and reduced task cost by up to about 50% compared with an agent without that graph. The company also gave an example in which a firm-knowledge task produced a 0% rubric result from raw files but 75% with the context graph, because the graph helped identify the relevant matter and executed credit and security agreements more quickly. These figures are company-reported research results, not independently verified production outcomes.
Harvey’s post also said that after the general availability of user memory last month, it surveyed users and found that 93% of respondents said memory makes Harvey more useful and 50% said they now repeat less context. The company framed the graph work as part of a broader investment in memory and said future additions will include precedent flags, cite checking for U.S. case law, law-evolution tracking, a new research surface and support for firm-defined legal ontologies. It said more details on those additions will follow, so a separate product launch date is not established in the post.
Be the first to comment