How memory works

These guides describe the current version of Space, which you can use today at app.getspace.sh.

Memory lets your agent save useful information and find it again in a later session.

Start with one preference

Imagine telling your agent that you prefer Python for scripting. You ask it to remember that preference, and it sends Space a statement to save: “The user prefers Python for scripting.”

That saved statement is called a claim. A memory can contain several claims, just as a note can contain several separate points.

Find it when it becomes useful

In a later session, your agent needs to write a script. It can ask Space which language you prefer. Space searches the saved information and checks possible matches before returning context the agent can use.

Finding saved information is called retrieval, or recall. The preference must have been saved, and the agent must request it. Connecting an agent does not automatically import your conversations.

Learn one concept at a time

Follow the same Python preference through these four short guides. Each introduces the words you need for the next step.

1. Memories, claims & sourcesWhat is saved, and where did it come from?
2. Saving a memoryFollow one sentence into a searchable memory.
3. Finding a memoryUnderstand meaning, vectors, cues, and indexes.
4. Choosing useful resultsSee how possible matches become returned evidence.

Inspect and correct what is remembered

A saved statement can be mistaken or become outdated. If you switch your preference to JavaScript, you can correct the claim. If it came from a document, its source link can help you check what was originally said.

Read Memory & sources for inspecting and changing saved context. Removing a claim from active recall does not erase the same information from source documents, retained revisions, or backups.

Understand what search can see

The saved text is encrypted. Search also uses readable numerical representations of meaning and fingerprints for matching. Those representations can reveal relationships and support guesses about the text.

The authorized service can read content, and configured model providers may receive text to process. See Privacy & data for the storage and processing boundaries.