legaltech.fyi Beta

The Op-Ed

Your New AI Assistant Has a Memory. That Is a Malpractice Trap.

Persistent context across Word and Outlook solves your prompt fatigue, but it quietly shreds client confidentiality if you do not clear the cache.

The legaltech.fyi editorial desk · 2026-08-20 ·3 min read

An AI-assisted editorial, reviewed by a human before publishing. It reasons over our own tracker data (and cited context) — a point of view, not legal advice.

What happens to your duty of confidentiality when your software remembers how you write?

If you ask the vendors, they will tell you that persistent context is the holy grail of legal productivity. Harvey announced Harvey II this week, rolling out a feature called "Memory" designed to retain an individual attorney's drafting quirks, citation habits, and workflow patterns across Microsoft Word and Outlook. Welcome.AI reports that the vendor spent months on listening tours with global firms to eliminate the friction of re-typing preferences into every fresh window. On paper, it sounds like an overdue relief for lawyers exhausted by writing the same prompt instructions twice a day.

If an AI assistant cannot forget where it learned its latest trick, it has no business sitting in your client's war room.

The Illusion of Harmless Preferences

The problem is that a lawyer's drafting preference is rarely abstract. We do not write in a vacuum. A preference for how to structure an indemnification carve-out or phrase a non-compete extension is shaped directly by the transaction sitting on the screen right now. When an AI tool retains that context to apply it to your next email or draft contract, the wall between Client A's proprietary posture and Client B's routine matter quietly dissolves.

This is not a theoretical software glitch; it is an old professional ethics breakdown wearing shiny new code. Writing on Law.com, Chad Ergun recently made the compelling case for zero data retention as a baseline requirement for law firm AI, arguing that retaining prompts and outputs threatens both competitive intelligence and attorney-client privilege. Yet while the market debates how long cloud servers hold onto raw files, vendors are actively selling features that harvest client-specific context and store it as a permanent profile on the associate's desktop.

Big Firm Scale Meets Small Error Margins

Consider the exposure scale. Davis Wright Tremaine LLP recently announced a firmwide rollout of Harvey alongside Microsoft Copilot and Microsoft's AI Frontier Suite, aiming for a 90 percent adoption rate across more than 600 lawyers and staff. When hundreds of timekeepers at a single firm use persistent tools every hour, a shared or persistent memory cache becomes a massive, unmonitored pool of cross-contaminated client data.

Meanwhile, general counsel are already tightening the vise. Our verified reporting shows corporate legal departments are actively moving past basic pilots to demand formal governance over autonomous AI, prioritizing vendor contract addenda, liability allocation, and strict compliance controls. If your firm promises an enterprise client that its trade secrets will stay strictly segregated, but your senior associate's AI co-pilot is silently carrying context from last week's hostile takeover bid into this morning's licensing deal, your security assurances are useless.

What to Do on Monday Morning

Fixing this does not require a year-long IT committee or another vague internal policy statement.

First, audit every AI integration currently running inside your firm's Microsoft Word and Outlook environments. Disable automatic cross-session context saving for any matter involving sensitive intellectual property, pending M&A, or regulatory investigations.

Second, treat AI memory clear-down as an ethical requirement on par with checking for conflicts. When an attorney closes out a matter or switches teams, clearing the tool's persistent memory cache must become a mandatory, logged closing step.

If an AI assistant cannot forget where it learned its latest trick, it has no business sitting in your client's war room.

From the archive

money 2026-08-17 · 3 min read
The $1.5 Billion AI Copyright Tax Will Kill Legal Tech Innovation

Anthropic’s record settlement draws a fatal distinction that prices boutique startup models out of the courtroom.

Read the op-ed →
practice 2026-08-12 · 4 min read
AI research is useful only after the lawyer verifies it

A new comparison study should end the sales pitch that legal AI replaces research instead of sitting in front of it.

Read the op-ed →
money 2026-08-05 · 3 min read
AI in law will reward proof, not enthusiasm

The firms that win will be the ones that can show where AI saves time, where it creates risk, and how that changes pricing.

Read the op-ed →
practice 2026-08-05 · 4 min read
UP’s AI ban gets the training question backwards

A first-year ban on generative AI may look cautious, but lawyers will not be entering a profession that lets them avoid the tool, only one that punishes them for trusting it blindly.

Read the op-ed →
courts 2026-08-05 · 5 min read
Courts are done treating fake AI citations as a slap on the wrist

The new Uprise sanction matters because it looks less like an embarrassment and more like the start of a repeatable discipline model for lawyers who file AI output without checking it.

Read the op-ed →
practice 2026-08-05 · 4 min read
Law firms should make AI training mandatory, not optional

Pinsent Masons’ safeguards, a wave of firm-side AI promotion, and recent sanctions all point to the same fix: supervised training has to become part of ordinary professional duty.

Read the op-ed →
courts 2026-08-04 · 3 min read
AI citation sanctions need to hurt more

California’s latest $10,000 fine shows the fake-citation problem is no longer a one-off embarrassment; it is a repeated professional failure that courts should punish publicly and hard.

Read the op-ed →
risk 2026-08-04 · 4 min read
AI filings are getting lawyers disciplined for a reason

The lesson from this week’s sanction cluster is simple: if AI drafts it, somebody must verify it before it leaves the building.

Read the op-ed →
courts 2026-08-04 · 4 min read
AI citation sanctions are becoming malpractice by another name

The courts are no longer treating fake AI citations as an embarrassment. They are building a repeatable sanctions routine, and lawyers who keep filing unverified output are doing it on purpose.

Read the op-ed →
practice 2026-08-01 · 4 min read
Texas just made the AI privilege fight about proof, not panic

A Texas Business Court minute entry did not bless ChatGPT as magic. It treated the chat log like work product, which leaves lawyers with the old questions: what was shared, who saw it, and whether anyone preserved the record.

Read the op-ed →
Get the op-ed in your inbox
Our take on law and AI, plus what actually moved — on the weekdays you choose.