Harvey II adds memory to legal AI after 3 acquisitions

Harvey, the four-year-old legal AI startup, launched Harvey II on Tuesday with Memory at its core, a capability that could deepen law firm dependence on its application layer as foundation models commoditize.

"Memory is all about the ability to learn your preferences — your choice of words, your drafting style, the structure of the work. If Harvey then makes a judgment call it can give answers based on those tailored preferences," Anique Drumright, chief product officer at Harvey, said.

The refresh introduces Spaces organized around matters or projects, consolidating documents, tasks, permissions, and work history in one place. AI agents now start with the context of a matter and user preferences already loaded, rather than beginning from scratch. The company has also been on an acquisition streak, buying product demo startup Hexus in January, AI-powered data integration platform Lume in March, and investment platform Benchmark in July.

The strategy places Harvey's application layer as the differentiator against large language model providers that can roll out increasingly capable models but lack the workflow integration and personalization Harvey is building. The more lawyers use the platform, the more tailored the output becomes, creating switching costs that could cement Harvey's position in the legal AI market.

Memory Extends Beyond Individual Users

Drumright said Harvey is working toward organizational memory — learning from what a law firm's lawyers and their clients want — while protecting ethical walls between matters. Client preferences on specific matters will connect to the context of each engagement, which Harvey links to its Shared Spaces collaborative forum between clients and law firms.

The approach mirrors a high-end hotel that learns guest preferences over time, according to Artificial Lawyer. Each interaction — whether drafting a contract, reviewing a brief, or communicating with a client — feeds back into the system, making subsequent outputs more aligned with the lawyer's working style. Memory has become a defining feature of AI assistants since early 2024, when major chatbot providers began rolling out persistent context capabilities, and Harvey's move to embed memory at the application layer reflects a broader shift toward specialized AI tools that understand domain-specific workflows.

Legal AI Competition Intensifies

DeepJudge, a legal AI search startup, unveiled an Agent Handoff Protocol this month that maintains context across AI platforms, with Harvey and Thomson Reuters already on track to implement it. Thomson Reuters, which trades on the NYSE under TRI, has been building its own AI-powered legal tools, including CoCounsel, making it a direct competitor to Harvey in the enterprise legal market.

Law firms are already integrating Harvey into broader AI stacks. ArentFox Schiff launched FoxAI this week, a framework that pulls together proprietary tools and licensed third-party AI platforms including Harvey, Microsoft Copilot, and DeepJudge to speed up client work.

The memory feature also addresses a practical pain point: lawyers who switch between drafting tools, email, and AI agents lose context at every transition. Harvey II's memory carries preferences across all three surfaces, reducing the need to re-explain drafting conventions or matter background. Harvey also launched a legal engineering certification course this week, a move aimed at building a talent pipeline of lawyers who can work effectively with AI tools.

Harvey remains privately held, but the legal AI market is attracting significant capital. The company's strategy of building deep workflow integration and personalization creates a moat that pure-play LLM providers like OpenAI, Google, and Meta cannot easily replicate, since they lack access to law firm data and workflow patterns. For publicly traded legal tech companies like Thomson Reuters, Harvey's continued product velocity raises the bar for AI feature development.

This article is for informational purposes only and does not constitute investment advice.