Choosing useful results

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.

A possible match still needs to be checked before it becomes useful context.

Candidates are possible matches

A search might find “The user prefers Python for scripting.” It might also find a saved note about a Python course. Both discuss Python, but only the first states the user’s preference.

A candidate is an item found for further consideration. Finding it does not mean it should be returned as evidence.

Ranking puts candidates in order

For meaning search, Space initially orders eligible claims by how similar their vectors are to the question’s vector. That ordering is called ranking. Search considers a limited set of candidates for the later checks.

The evidence check asks whether the text helps

In the default memory search, a model reads the retrieved text and assesses whether it supplies evidence for the question. A statement about taking a Python course does not, by itself, establish a preference for Python.

“Selected” means chosen at a particular step

When we say “selected candidates,” we mean the items chosen to examine. When we say “selected results,” we mean the items chosen to return. It describes a choice, not a separate index or a special kind of memory.

A claim can be saved and searchable without being selected for this question. Your preference for short explanations can stay in memory even when the agent only needs your preferred scripting language.

An empty result is possible

If the retrieved evidence cannot answer the question, the search can return no answer evidence. This does not prove the information is absent from every saved record. Retrieval and assessment both have limits.

The agent uses returned context to help answer you. It should still inspect sources and consider whether the information has changed.

Keep exploring

Read Agent memory explained for the broader relationship between saved memory, the current conversation, and model learning. For a worked example of checking evidence, read An AI memory should tell you where it came from.