How memory works

Memory is everything the assistant can find and use when answering you. It lives in the workspace on your devices.

What it consists of

Private chats and documents are not part of memory. See Private chats.

How memory gets into an answer

Automatically. Before answering, MindDuck searches memory for what relates to the question and passes the fragments it found to the model. Most often this happens when the question itself hints at memory: "what did we discuss about…", "do you remember…", "what are my preferences…". Assistant has an experimental Search memory for every text message switch: with it, search runs before every message. If Smart memory selection is on, the Jev model on OpenRouter picks from a wide list of results up to five fragments that help answer, plus up to four "associations": fragments on the same subject from other conversations that don't answer directly but may remind of something. The model receives them separately and uses them only if they really help. Under the answer they are shown in the memory fragments section marked Possibly related. If there is a delay or an error, the regular selection is used.

When the model decides. The model can itself call memory search, open a document or look up a fact if it needs to. All such actions are visible in the App actions block under the answer. With Smart memory selection, the model's search also goes through Jev's selection.

Facts are used in every conversation where they fit: global facts in any chat, project facts only in that project's chats.

Memory search runs on the device. Only the fragments found go to the model, not the whole base. With Smart memory selection, short fragments from the result list (up to 300) also go to OpenRouter for the selection.

Nothing is remembered in secret

• The assistant can only suggest a fact. A fact gets into memory only when you confirm it.
• Confirmed facts don't disappear on their own. They can be corrected, stopped being used or deleted, but you are the one who does it.
• Every fact shows where it came from: which message and when.

Whether the assistant may suggest facts at all is up to you: Settings → Assistant → Let the assistant suggest facts.

Sections of the Memory screen

Search: search across all of memory. Documents: the document library. Facts: saved facts. Projects: projects. Chats: all conversations. The trash button is Trash.

Memory size on the Memory screen shows the amount of memory in words and approximate tokens.