How memory works
Space is in development. These guides describe the current development version and the intended workflow. Public sign-ups and hosted agent connections are not available yet.
Persistent memory gives your next agent session useful context from earlier work. Space keeps that context in a workspace you can inspect and correct.
What is AI agent memory?
Memory is saved context an agent can retrieve later: a decision, a preference, a fact about a project, or the next step in a piece of work. For example, you might ask an agent to remember why you chose a particular design. In a later session, the agent can retrieve that decision instead of asking you to explain it again.
Save context deliberately
An authorized agent can write individual memory claims directly. It can also submit text or a saved document revision for a background processor to extract claims. Accepted claims become available without a human approval step. Background extraction takes time; submitting work does not mean it has finished.
Connecting an agent does not import earlier conversations. Your agent must explicitly send the context it wants to save. You can inspect, correct, or delete remembered claims in the development app.
Keep documents, tasks, and memory together
A document preserves the detailed work. A task records something to do. Memory captures reusable context. A claim derived from a saved document can refer to the source revision, so you can understand where it came from. Directly written claims can have different evidence; not every claim has a document behind it.
Find useful context in a later session
A connected agent asks Space for context through the Model Context Protocol (MCP). Space searches the authorized workspace using meaning and matching cues, ranks candidates, and returns a bounded set of results. The current development implementation ranks eligible memory within the selected space.
Retrieval is not a guarantee of truth or relevance. An incomplete or incorrect memory can still be retrieved. The agent should check the evidence, explain uncertainty, and distinguish saved facts from its own conclusions. Read how agent connections work for the setup and access steps.
Understand the privacy boundary
Claim text is stored in encrypted revisions. Search also uses readable embeddings and content or cue hashes. These search representations can reveal relationships and support guesses about content, so encryption does not make the entire database opaque.
The authorized service processes readable content, and configured model providers can receive selected text for extraction, embedding, or relevance checks. Your connected agent receives the results it is authorized to read. See Privacy & data before deciding what to store.
Keep memory useful as work changes
A source can change after a claim was extracted. Inspect the source revision and update or remove stale memory when needed. Removing a claim from recall is separate from deleting source documents, retained revisions, or backup copies.
Space is still in development. These guides explain the current design and implementation; public hosted access and a supported self-hosted release are not available yet.