Memory
Understand saved facts, conversation recall, and how memory differs between Chat and Code.
Memory across the workspace
Memory is a shared concept with different storage and access on each surface. Chat has account-backed saved facts and optional conversation recall. Code retains local task history and runtime context; linked Chat recall requires an explicit project link and a supported installation. Files can provide document context to Chat, but uploading a file is not the same as saving a memory.
| Surface | Memory behavior |
|---|---|
| Chat on PC, web, and mobile | Uses the signed-in account's memory controls, project scope, and agent policy |
| Desktop Code | Uses local task and provider context; optional linked Chat recall is described in Memory in Code |
| Desktop Personalization | A separate local preference list for qualifying general conversations; not automatic repository-task memory |
| Mobile Remote | Uses the paired desktop's task context and permissions |
View diagram text
flowchart TD accTitle: How Chat memory informs a reply accDescr: Saving durable facts and indexing past conversations are separate optional paths. Relevant allowed context is selected for the current project and agent before a reply. A[Conversation] --> B[Dynamic Memory enabled] B --> C[Global or project facts] A --> D[Chat History Memory enabled] D --> E[Searchable past conversations] C --> F[Saved-memory use and scope checks] E --> G[History recall and scope checks] F --> H[Relevant context allowed for this agent] G --> H H --> I[Current reply]
Saving facts, using saved facts, and indexing conversation history are separate controls. The diagram shows available paths, not a promise that every message is saved or every reply searches history. Review and correct facts in the memory list.
Choose what Chat can remember
Settings > Memory separates three decisions. Dynamic Memory lets the model save and maintain durable facts stated in chat. Use saved memories lets relevant facts inform replies. Chat History Memory lets finished text turns become searchable for future conversation context. Saving and using a saved memory are separate controls, and chats remain in the sidebar when chat-history indexing is off.
The Projects section lists a Dynamic memory for each project. Turning on a project switch allows facts from that project’s chats to be saved in that project’s scope. Global facts and project facts stay distinguishable in the memory list. A project can also have its own Dynamic memory switch in Projects.
Review and correct facts
The Memories section can filter by Scope and Tag. Each fact shows its text, Global or project scope, and tags. The row actions are Edit memory, Mark verified, and Delete memory. Delete asks Forget this memory? before removing the fact. Tags keep related facts grouped; a memory can carry up to five tags, and New tag adds a label for later filtering.
When the assistant saves or updates a fact during a chat, a memory chip appears in the transcript. The chip can expose Undo for an add or update, so a mistaken change can be reversed close to where it happened. Reviewing the list remains the reliable way to inspect older facts.
Manage chat history
The Chat History section shows an Indexed chunks count and explains that finished text turns are divided for conversation search. Wipe deletes the indexed chunks and stops indexing until Chat History Memory is enabled again; the original chats stay in the sidebar. Changing the embedding model reprocesses memory, history, and file chunks in the background, so search may be incomplete while that work finishes.
Memory is scoped to the current project and global context when a project chat is active. Agent memory follows the agent’s Memory access choice: Agents explains the Read-only and Off options. The Search article separates sidebar title search from memory-assisted context retrieval.