Harvey Graph connects legal matters, documents and case law before agents begin work

Harvey’s new context layer maps relationships within matters and across US case law, reducing the repeated briefing behind legal-agent tasks.

Harvey has introduced a context layer that gives legal agents a structured view of the material surrounding an assignment. In its [19 October announcement](https://www.harvey.ai/blog/improving-agent-accuracy-with-harvey-graph), the company describes Harvey Graph as a combination of a US case-law knowledge graph and private context graphs within individual Spaces.

The current system connects legal authorities and citations, while the matter-specific layer maps documents and information from Vaults and integrations. Harvey says it follows existing Space governance and security settings. Features planned for subsequent months include precedent flags, citation checking, tracking changes in legal authority and customer-defined legal ontologies.

The relationship is part of the answer

A contract question often depends on more than finding a paragraph containing the right words. A signed agreement can supersede a draft; an amendment can change one obligation while leaving the rest intact; a side letter can qualify how a provision applies. Those relationships determine which passage answers the question.

A conventional similarity search can return all three documents without resolving their order. A graph creates an additional place to represent the connection: this document amends that agreement, this entity is a party to that transaction, or this opinion relies on that earlier case. The agent can then use that structure when choosing its next search or preparing an answer.

That approach can make repeated work more efficient. If the relationship between two agreements has already been established, a new task need not infer it again from filenames and scattered references. The engineering challenge shifts toward maintaining the relationship as new documents arrive and as a matter changes.

Knowledge and matter context serve different purposes

The two graph types address different questions. A legal-knowledge graph helps navigate authority: which opinions cite an earlier decision, and how does a line of reasoning develop? A matter-context graph helps navigate a particular client assignment: which document belongs to this transaction, which version is executed, and what does the team already know?

Bringing those perspectives together can support a more useful first draft. Consider a refinancing review. The agent needs the correct executed security agreement from the client's previous transaction and relevant current authority for the new work. Better retrieval of one does not automatically solve the other. Keeping the two forms of context distinct makes the research path easier to understand.

A different use for organisational memory

The announcement also links graph development with Harvey's investment in memory. This creates two complementary kinds of continuity: preferences about how a professional works, and information about what a particular matter contains.

The distinction matters operationally. A preferred drafting style can travel with a user, while a client's deal documents belong within a defined matter and its access rules. A useful product can remember both without treating them as interchangeable.

For legal teams, the immediate development is therefore a change in how assignments begin. Instead of repeatedly assembling context inside each prompt, an agent can start from an organised view of the work already held in the platform. Reviewing the answer still involves the underlying documents and authorities, but the route to those materials can become more direct. Harvey's release puts that context-building layer at the centre of its legal-agent strategy.

Sources

- [Improving Agent Accuracy With Harvey Graph](https://www.harvey.ai/blog/improving-agent-accuracy-with-harvey-graph) — Harvey