Memory
Assistant remembers facts about you from your messages: preferences, names, projects, habits. A fact is a short pair such as "favourite editor: Neovim". The model reads the facts through the Knowledge tool when they fit the question. Settings → Knowledge Base holds the facts, the models and the logs.
Set up#
Memory needs an embedding model, a small model that turns text into numbers. Select one under Settings → Knowledge Base → Models. The list shows the embedding models of your engines; the setup wizard downloads one with the starter model. Without it, memory is off.
Optionally select a chat model for the extraction. Empty uses the model of each conversation, which avoids a model swap.
How facts are collected#
Assistant collects facts in the background, when you have not written for a few minutes (Knowledge idle minutes under Settings → Engines → Server, 5 by default) and nothing else runs:
- It reads every new message of yours and writes down the facts.
- It computes the embedding of each new fact. A fact that says almost the same as an old one is compared with it.
- It keeps, updates or adds the fact.
A fact shows an Indexing badge until step 2 is done. Run Now starts the collection at once.
Manage facts#
- Discovered Facts: search, edit or delete a fact. Add one by hand.
- Request Knowledge Rescan deletes every fact and reads all your messages again.
- Execution Logs: the last messages checked, with their result.
Deleting a conversation deletes the facts it produced. Temporary chats and group conversations never feed memory.