Memory
File-backed managed memory that gives the agent durable, reviewable context across conversations.
How it works
Memory is stored as markdown files under memory/ in the data directory (and per-workspace). The agent writes to it with the update_memory tool, and relevant snippets are recalled automatically into later conversations — no manual attach needed. Because it's files, memory is inspectable, searchable, and editable by hand.
| Scope | Purpose | Default size limit |
|---|---|---|
| User | Facts about you — role, preferences, how you like to work. | 2,000 characters |
| Workspace | Facts about a specific project. | 3,000 characters |
Background review
A background pass periodically reviews recent conversation turns and folds anything worth keeping into the vault — deduplicating, trimming, and linking related notes. Defaults:
| Setting | Default |
|---|---|
enabled | on |
tool_enabled (agent can write directly) | on |
background_review_enabled | on |
review_interval_turns | 10 |
background_review.model | follows the chat model unless set |
Managing memory
The Memory tab in Settings lets you browse the memory files, search their contents, adjust size limits and review cadence, and choose the review model. Smarter recall and upkeep improve which snippets surface in a given conversation.