Finding a memory
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.
Search can find a saved statement even when your question uses different words.
Start with meaning
Later, your agent asks: “Which language does the user prefer for scripts?” The question and the claim use different words, but their meanings are related.
A vector is a list of numbers
An embedding model converts text into a list of numbers called a vector, also called an embedding. These numbers allow the system to compare texts for similarity in meaning.
Picture a map where similar ideas tend to sit near each other. Space compares the question’s vector with saved claim vectors to find possible matches. The numbers do not establish whether a statement is true or whether it answers the question.
A cue is a label that connects claims
During extraction, the processor can also generate short labels called cues. Our Python claim might receive a label such as “scripting.” This is an example, not a guaranteed label.
If another claim has that same cue, Space can find it as a related memory. This lookup matches stored cue labels exactly. It is different from searching for the word “scripting” anywhere in the claim text.
Related memories are optional and returned separately from ranked answers. Directly written claims currently have no generated cues, so meaning search is still useful on its own.
What does “one claim, two indexes” mean?
A database index is a lookup structure that helps with a particular kind of search. Space has two of these structures over the same claim records: one for vector similarity, and one for matching cue labels. They are not two copies of the claim text.
Under the hood: HNSW, GIN, and the default search
- HNSW: a shortcut for vector similarity
- This index connects nearby vectors so a search can explore promising matches without comparing every vector. That makes it approximate: it can miss some of the nearest matches.
- GIN: a lookup for shared cue labels
- This index supports looking up claims that share a cue. The database stores hashes of the labels—fingerprints used for matching—rather than the readable labels themselves.
- Default search: compare eligible vectors directly
- The current development version compares all claim vectors within the authorized Space and requested embedding model, then takes the nearest matches. This is exact vector ranking. It bypasses the HNSW shortcut; HNSW remains available for explicit evaluation.
“Exact” describes the vector comparison. It does not guarantee a useful or correct answer. The next page explains how Space checks possible matches.
Keep these words separate: an index is a lookup structure; the default is the method normally used; selected results are the items chosen at a search step.