Next Gen ‘Harvey II’ Launches with Memory at Its Core

Harvey II Bets Your Firm Wants an AI That Remembers You — Not One That Shows Its Work

Harvey has rolled out a major platform refresh, branded Harvey II, built around a new "Memory" capability. The update was announced by chief product officer Anique Drumright and first reported by Artificial Lawyer in August 2026. Alongside Memory, Harvey II adds preference setting, context sharing between lawyer and client, matter-centric "Spaces" that bundle documents, tasks, permissions and history, and agents that start work already loaded with a matter's context and a lawyer's known preferences. Drumright frames it as a system that, in her words, "gets better and better the more you use it."

What does "Memory" actually change day to day?

Harvey now retains a lawyer's drafting style, word choices and structural preferences across Word, Outlook and its own agents, and applies them automatically to future output.

That's a shift from a tool you prompt each time to one that carries state between sessions. Spaces extend the same idea to a matter or project, holding documents, tasks and permissions together so agents can pick up "substantive legal work" with context already in place, according to the company's own description. It's a plausible productivity story — less re-explaining, less boilerplate correction — but it's also, so far, a vendor's account of its own feature, not an independently measured one.

Does a system that learns from your matters raise privilege or wall problems?

Yes, in principle — and Harvey itself flags this rather than ignoring it, which is the right instinct but not a substitute for detail.

Drumright told Artificial Lawyer the update includes a "major focus on context and Memory, and how we can leverage Memory and protect ethical walls." That's a real tension worth naming: a feature designed to carry a lawyer's preferences and a client's preferences across matters is, by construction, a feature that moves information across boundaries. The public reporting doesn't say how walls are technically enforced — whether Memory is scoped per matter, per client, per ethical-wall group, or globally per firm — and that's the detail a buyer needs before rollout, not after.

How does this sit next to evidence-grounded AI tools firms already use?

Harvey II optimizes for fluent, personalized drafting; evidentiary platforms optimize for provable, cited answers — different jobs, different risk profiles.

Harvey's pitch is that the application layer — your precedents, your working style, your client's preferences — becomes something no general-purpose LLM can replicate on its own. That's a strength for drafting and correspondence. But for discovery and fact investigation, the value a firm needs isn't personalization, it's traceability: every answer tied to a specific exhibit, with chain of custody intact, so opposing counsel and the court can check the work. A memory layer that quietly shapes phrasing and judgment calls is a different kind of system from one built so every output is auditable against source documents. Firms running both will want to know which category a given output falls into before it goes in a filing.

What should buyers ask their vendor before adopting this?

Ask exactly what Memory stores, where, for how long, whether it's scoped by matter or client, and whether any of it touches model training.

Also worth asking: does Memory run single-tenant or shared infrastructure, can a client demand their preference data be purged at matter close, and who audits that ethical walls actually hold once Memory spans multiple engagements for the same lawyer.

Frequently asked questions

Is Harvey II a new product or an update?

It's described by Harvey as a major refresh of the existing platform, not a separate product, centered on the new Memory capability.

Does Memory work across Harvey's whole product surface?

According to the company, preferences carry across Harvey, Word, Outlook and Harvey's agents, not just within the core chat interface.

Has any third party verified the "gets better the more you use it" claim?

No independent measurement has been reported; the claim currently rests on Drumright's own description to Artificial Lawyer.

Source: Artificial Lawyer, "Next Gen 'Harvey II' Launches with Memory at Its Core."