Harvey Launches Harvey II With In-House Tenet Model and Persistent Work Memory
The AMW Read
Harvey, a named case-study player in Legal AI, shifts from third-party foundation models to an in-house model and adds a persistent memory layer, updating the segment's competitive baseline without resolving an open debate.
Harvey Launches Harvey II With In-House Tenet Model and Persistent Work Memory
Harvey, the US legal AI startup, introduced its Harvey II platform on August 18, built around two new components: Tenet, the company's first in-house legal-specific model, and a "Memory" feature that retains how individual lawyers prefer to edit documents and format output. Tenet is built on Moonshot AI's open-weight Kimi K3 model, fine-tuned using synthetic legal disputes and case files created by practicing lawyers. Harvey had previously routed legal tasks through third-party models from OpenAI and Anthropic; owning Tenet is intended to cut those inference costs. Harvey has raised $1.21B to date, per the AI Market Watch index (tracked among roughly 5,000 companies — coverage, not a census).
Memory extends personalization beyond the single-user chatbot pattern that OpenAI, Anthropic, Google, and Microsoft have already built into general-purpose assistants. Harvey is applying it to matter, client, and case-level context, with a stated goal of eventual organization-wide context — distinct from Tenet's legal knowledge, since memory tracks how a specific firm or lawyer works rather than what the law says. Harvey's chief product officer told the Wall Street Journal the goal is moving from an AI that answers individual questions to one that understands the context of ongoing work. The move signals that competition in domain-specific enterprise AI is expanding past fine-tuning and retrieval-augmented generation into a persistent-context layer that rivals will now need to match.
For law firms evaluating legal AI vendors, switching costs may rise as memory accumulates firm- and matter-specific context that isn't portable to a competing platform. Investors assessing legal AI should distinguish model-quality claims from this kind of workflow-embedded data advantage, since the latter compounds with usage in a way a fine-tuned model alone does not.
