Knowledge

When a new chat doesn't remember

You saved something to a Brain in one chat, opened a new chat, asked about it, and the assistant had no idea. A Brain works only in the chats it is attached to. There is no default Brain for a workspace, a folder or a person, so every new chat starts without one.

To check whether a chat has a Brain, look at the Brain chip among the message box's controls. In a new chat the chip reads No memory beside a small dashed circle, and on a narrow screen you see only the circle.

Attaching a Brain

Click the chip. A panel titled Choose a brain for this chat lists No memory first, then every Brain in the current workspace with its entry count and description.

Any workspace member can attach, swap or remove a Brain. Owners and admins also get New brain at the bottom. Its form's Create brain button makes the Brain and picks it for this chat.

If the workspace has no Brains yet, the panel says Create your first brain instead. Only owners and admins get its Create a brain button, so a member has to ask one of them.

Creating a Brain needs a plan that includes Brains and room under its Brain limit. When the workspace is set to Owner Credit the owner's plan counts, and under Individual Credit your own does. When the plan blocks it, New brain opens Upgrade Required, which says Your plan does not include Brains or states the workspace's limit.

In a new chat

Pick a Brain before your first message, and sending that message attaches it. Right after that send, the chip dims until the chat exists, and its tooltip reads Waiting for the chat to be created… Reload the page before sending and the pick is gone.

In a chat that already has messages

The chip spins until Lexika saves the change, and the send button stays greyed out until then. Press Enter on a physical keyboard in that moment and a toast reads The Brain attachment is still updating. Try again in a moment. The new Brain applies from your next message. An answer already being written keeps the Brain it started with.

To take a Brain off, pick No memory. Everyone with the chat open sees the chip change and a toast such as Brain 'Launch plan' detached.

One Brain per chat, shared by everyone in it

A chat holds one Brain at a time. Picking another replaces the first. When several people in the workspace send messages in the same chat, every message uses that chat's Brain, whoever attached it. The Knowledge Base article on how things get into a Brain explains who may save to it.

Branch in new chat, Regenerate and Edit message are icon buttons on a message. On a computer, hover one to see its name.

When you	The chat's Brain
Branch in new chat from an answer	The branch keeps it
Regenerate an answer or edit a message	The new answer uses the Brain attached now
Use Fork conversation on a shared link	The copy starts at No memory
Delete the conversation	The Brain is detached, and memories saved from that chat stay in it

When an owner or admin deletes a Brain, every chat using it returns to No memory, and anyone with one open sees The Brain 'Launch plan' was deleted and detached from this conversation. Deleting a Brain also deletes all its memories for good.

What the assistant sees, and when it looks

With a Brain attached and a model that can read it, every answer starts from a short profile. It holds the Brain's name, its entry count, the description written under What belongs in it when the Brain was created or edited, and the summary the Brain's page shows under What's inside. The memories themselves are not in it.

Lexika tells the assistant to search when a question may touch saved knowledge. The assistant decides each time whether to search. It can use the Brain at most six times in one answer, so ask about one topic at a time when you need many memories. A later answer sees what earlier answers said, but not the memories they read, so it has to search again to get the details.

Lexika rewrites What's inside from the newest memory titles about five minutes after the assistant saves or removes something. Your own edits and deletions on the Brain's page don't trigger a rewrite, and an emptied Brain keeps its old summary. So the assistant can mention a memory that is gone, and a search won't find it.

Lookups add steps of work for the model, and the message's final charge counts every step, so an answer that searches the Brain usually costs more. The Knowledge Base article on what a message costs explains the final charge.

Checking what the assistant looked up

Each lookup leaves a row above the answer text, among the assistant's other steps, in the order it happened.

• Searched the Brain, then the words it searched for, and a count such as 6 memories or no matches.
• Read Brain entries means it opened whole memories.
• Followed Brain links, then a memory's title, means it looked at the memories linked to that one.

Click a row to open it. A search lists each memory under Matched in this Brain with its title, a percentage and the start of its text.

The percentage measures how close a memory's meaning is to the search words, not whether the memory is correct or the answer used it. A search returns its closest memories, six by default and never more than ten, even when none fits well. A memory found by matching words instead of meaning shows no percentage. Lexika asks the assistant to name the memories it relied on, so look for them in the answer text.

An answer with none of these rows read no memories. Whatever it knew about your saved work came from the profile and this chat's earlier messages. If it missed something you know is saved, ask it to search the Brain, then open the row to see what came back.

When an attached Brain goes unread

The chip shows the Brain's name whichever model you pick, but only models that can use tools can read a Brain. Perplexity models and image models can't, and the model selector doesn't say which others can't. Such a text model knows only the Brain's name, and if you ask about remembering, it tells you to switch models. An image model doesn't know about the Brain at all. If you asked for a search and no Searched the Brain row appears, switch to another model and ask again.

Council and Agentic Research, the Mini Apps, have no Brain and no Brain chip.