AI memory arrived as a headline feature on 18 August 2026, when the legal platform Harvey relaunched with a system that learns how each lawyer works and carries those preferences into Word and Outlook. Your firm is unlikely to have bought that product. The reason to read on anyway is that memory is arriving in the general assistants you already pay for, and it lands on the same duty in the same place, being the wall you keep between one client and another.
The supplier is rolling the feature out in stages. Personal memory comes first, holding drafting style and word choice for one fee earner. Shared memory across a matter space follows in the months ahead, and firm-wide memory sits at the end of the road. Users can see what the tool has remembered, change it or switch it off, and the supplier states that memory is never used to train models.
Where memory meets the wall
Personal memory of how you like a clause worded raises little. The question sharpens as memory widens. A store that learns from every matter it touches has read across your files, and the point of an information barrier is that some of those files stay out of reach of some of your people. Harvey has answered that in the enterprise way, syncing ethical walls from a firm's existing conflicts system, letting a firm set view, edit and run permissions separately for each organisation sharing a space, and giving administrators a timestamped audit log for access reviews.
Those are controls a two-hundred-partner firm asks for and pays for. The assistant on your desk has none of them. It has one memory, one user and no notion that the matter you worked on yesterday is one your colleague must never see.
What to ask before you switch it on
Put four questions to any supplier offering memory. Ask whether the memory belongs to the user, the matter or the firm, since only the first is safe without a conflicts check. Ask where the store sits and how long it lives. Ask whether closing a matter and deleting its documents also removes what the tool learned from them, because your retention schedule now has to reach a record nobody wrote. Ask, in writing, whether memory feeds model training.
Then decide who turns it on. Memory is a setting in most consumer assistants and it is often on by default, so a firm that has said nothing has already chosen. Add a line to your AI use policy naming who may enable memory and on what work. Tell your fee earners that anything they type into a remembering tool outlives the session. Then check the setting on the accounts your people already use, this week rather than at the next policy review.
The trade press reported the launch on 18 August 2026, and the supplier sets out the ethical wall, permission and audit controls in its own open monthly release note for August 2026.
If your firm is weighing a tool that remembers and wants the conflicts and retention questions settled before anyone switches it on, that is work we do with firms in an afternoon: talk it through with us.