Harvey announced on August 18 the launch of Harvey II, a product restructuring based on three pillars: persistent context per case file (“Spaces”), the lawyer’s personal memory that is transferred between Harvey, Word, and Outlook, and a proprietary language model—Harvey Tenet—specifically trained for legal reasoning. It is the company’s first proprietary model after four years of operating on generic third-party models.
The change in architecture addresses a structural limitation of the previous product: each task started from scratch. The lawyer had to provide new documents, explain the subject matter, and correct the outcome with each intervention, even within the same case file. Harvey II reverses this starting point—agents automatically inherit documents, reports, tasks, permits, and ethical boundaries from the case in which they are opened—which the company presents as a condition for delegating more complex legal work, not just one-off tasks.
Memory is being rolled out in three phases. The first—individual drafting and structuring preferences—is currently in early access. The second, planned for the coming months, will extend Memory to a shared workspace. The third, with no set date, will bring it to the organization level, with law firms and departments contributing their own workflows within the relevant confidentiality boundaries. Harvey anticipates general client access to Memory in the third quarter of this year.
Harvey’s statement clarifies two operational conditions not typically included in product announcements but relevant for those authorizing the use of these tools with client documentation: permissions and ethical safeguards for each case file are synchronized directly from the firm’s existing systems to the corresponding “Space,” and AI spending is tied to that specific case file, not an aggregated account, allowing for cost tracking by case and by client. Regarding Memory, Harvey adds that the lawyer can view, modify, or completely disable what the system retains, and that this information is not used to train the models.
The most promising development for the sector is Harvey Tenet. Until now, Harvey relied on general-purpose models—primarily from OpenAI, although it tested other providers. According to the company, Tenet is a front-end product in the main legal benchmarks, with an operating cost comparable to that of open-source models. This would allow agents to run continuously on each case without the cost associated with general-purpose models. This performance and cost comparison is a claim made by Harvey, with no independent verification available to date. The company also plans to build client-specific models based on its own work history, meaning that two firms using Harvey would end up with different models.
The decision comes at a time of direct competitive tension. Stockholm-based Legora closed a $550 million Series D funding round in 2016, led by Accel, with participation from Benchmark, Bessemer, General Catalyst, and ICONIQ, valuing the company at $5.55 billion. Legora builds its product on Anthropic’s Claude platform and has focused its growth in the United States, in contrast to Harvey’s European expansion.
Harvey, meanwhile, reached a valuation of $8 billion last fall, following a $160 million funding round led by Andreessen Horowitz that raised its 2025 funding to $760 million. The firm claims more than 1,000 clients in 60 countries, including most of the top ten U.S. law firms by revenue.
At Líder Legal, we see the relevant data for the Ibero-American market not in the valuation figure itself, but in the shift in purchasing criteria. While first-generation legal AI tools are evaluated based on research accuracy and drafting speed, Harvey II shifts the comparison to a different arena: which provider retains the context of a case file between sessions, who controls that context once stored, how is spending tracked per case, and what underlying model supports the reasoning? These are the questions that, in our view, will begin to appear in the vendor selection processes of law firms and legal departments that currently operate with general-purpose, prompt-based models, lacking a memory layer or integrated document governance.
