Retentive.
It grows with your matter, not your session.

"Let's continue where we left off."

I have seen that sentence typed into an AI drafting tool more times than I can count. The attorney is returning to a matter they opened three weeks earlier. There was a claim strategy discussion at the end of the last session; a decision about whether to pursue broad functional claiming or narrow structural claiming in independent claim 1; that they want to pick up and finalise.

The AI has no idea what they are talking about.

Not because it is not powerful enough. Because it was never built to remember. The session ended, the context window closed, and everything that was discussed; every decision made, every nuance explained; evaporated. The attorney is not continuing a conversation. They are starting a new one and hoping to reconstruct the old one from their own notes.

This is not a niche problem. It is the default state of almost every AI tool that touches legal work.

Why amnesia is a design choice, not a technical limitation

I want to be direct about something because I think it is often misunderstood. The fact that AI tools forget everything between sessions is not an unsolved technical problem. It is a product decision. Persistent, structured, matter-level memory requires storage design, retrieval architecture, and a thoughtful model of what is worth remembering and what is not. Those are hard problems; not impossible ones.

Most tools do not solve them because their initial use case is a single-session interaction. You have a question; the AI answers it. You want a draft; the AI generates it. That flow does not require memory. It works perfectly well as a stateless transaction.

Patent prosecution does not work that way. A matter that starts with an invention disclosure in January might be in prosecution in September. The office action response in October needs to be consistent with the claim strategy from January. The continuation application in February of next year needs to trace its lineage back to the original filing in a specific and defensible way. The entire arc of a patent matter is one continuous conversation; and a tool that forgets between sessions cannot participate in that conversation. It can only respond to isolated moments within it.

What gets lost when a tool forgets

The practical costs of session-scoped AI in patent work are specific and significant. They are worth naming clearly.

• Claim strategy drift. A decision made about claim scope in session one; for a specific reason, informed by a specific prior art concern; is not available in session two. The attorney either has to re-explain it or trust that the AI will independently reach the same conclusion. It usually doesn't.
• Figure inconsistency. Patent figures are numbered and cross-referenced throughout the application. When an attorney adds a figure in session three, the reference numeral assignments from sessions one and two are gone. The tool generates new ones. The application now has internal inconsistencies that require manual reconciliation.
• Repeated context rebuilding. Every session begins with the attorney spending time re-establishing who the inventor is, what the invention does, what the key claims are, what decisions have been made. That time is not recoverable. It is a fixed tax on every interaction with the tool.
• Prosecution history gaps. When the first office action arrives and the attorney goes back to the AI for a response strategy, the tool has no memory of why the claims were drafted the way they were. The reasoning that would justify the original scope is gone.

What Retentive means in eety

eety stores the full history of a matter as a set of structured, persistent artifacts. Not a chat log. Actual structured data that the system can reason from.

• Invention understanding as a living artifact. The model eety builds of an invention; its fifteen-dimension analysis of the inventive concept, components, novelty, and gaps; is stored and updated as the matter develops. When new information arrives, it is integrated. When the attorney clarifies something, the model is updated. The understanding grows with the matter.
• Figure memory. Every figure eety generates or processes is stored with its reference numeral assignments and its relationship to the claims. When new figures are added, they integrate into the existing numbering scheme. When figures are renumbered, the change propagates automatically through the specification.
• Section-level history. Every version of every section is stored with a record of what changed, when, and in response to what instruction. The attorney can see how the claims evolved from the first draft to the current state and understand the reason for each change.
• Claim architecture decisions. When the attorney approves a claim architecture; or modifies it; that decision is stored as part of the matter record. It is available in every subsequent session as the baseline from which new work proceeds.

The difference in what "returning to a matter" actually feels like

The practical test for retentiveness is simple. Open a matter you last worked on three weeks ago. What is the experience?

With a session-scoped tool: the attorney is back at the beginning. They know what was decided; the tool does not. They spend the first twenty minutes re-establishing context before any productive work can happen.

With eety: the matter is where they left it. The invention understanding is current. The figures are consistent. The claim architecture is intact. The attorney opens the matter and continues; not reconstructs.

That difference does not sound dramatic when you describe it. It is dramatic when you live it across forty active matters simultaneously.