Ask your memories, and see the segments that answer.
"How have I translated indemnity?" "Are my notice headings consistent in case?" Plain questions, answered from real searches of the memories this project reads, with where each answer came from.
Ask it in those words. The variant audit pulls every rendering of a term across your read-TMs and groups them by what you actually wrote. The drift check groups by shape rather than wording — length, ending, capitalisation — which is how you find the three segments that were translated on a different day.
Term, reverse, proximity, wildcard, fuzzy, Boolean and a cheap count check — seven ways into your memories, and the agent chooses between them the way you would. What comes back is a real search result, not a summary of one.
Fifty hits per call. A count check before any unfiltered query, refusing above two hundred. Those limits live in the retrieval tools themselves, not in the instructions given to the AI model, so they hold whatever the agent is asked.
The agent searches the memories the project already reads. Change that in project settings, as you always would: one list, the project's, and the agent searches exactly the memories on it.
The stand-ins that protect your client's material in translation apply to every search the agent runs, from the same rules: your question, the segments it reads, the names of the memories and documents they came from, and the domain, project and client names it is offered to narrow a search by. Every request the agent sends is read back before it goes, the whole conversation each time, and refused if a protected name is still in it.
Searching your memories this way needs an AI model that can work with tools over several turns. You choose it in the Agent Console, the same choice as for Agentic Editing: a wide one, across five providers, refreshed when the application starts. Add the provider's key in LLM Configuration and its models are there to choose.
When the search shows a term rendered three ways, name the one to keep. The agent finds the segments in this document that use the others and proposes the change in each, in the same queue and with the same per-segment review as any other edit; a locked segment is left as it is, and named.
