Skip to content
// Glossary · Agent memory

Semantic search

Also called: vector search, embedding search, retrieval by meaning

What is semantic search?

Semantic search retrieves results by meaning rather than by matching words. Text is converted into a vector embedding, and a query finds the entries whose embeddings sit closest to it. That is why a search for authentication flow can return a memory written about the login handshake, which shares no words with the query.

// In more depth

How retrieval by meaning works.

Three steps, and the second is where most of the quality lives.

01

Text becomes a vector

An embedding model turns each entry into a list of numbers positioned so that similar meanings land near each other. This happens once, when the memory is written.

02

A query becomes a vector too

The search text goes through the same model, then the store returns the nearest entries by distance. Nothing is compared as literal text at any point.

03

Ranking decides what surfaces

Nearness alone is not enough in a corpus that has been accumulating for a year. Recency, importance, and usage shape what actually comes back first.

// Why it matters

Why it matters for agent memory.

The alternative is requiring an agent to guess the exact words a different agent used months earlier.

Vocabulary stops mattering

Nobody has to remember the phrasing used when a memory was written, which is the failure mode keyword search hits immediately.

It works across authors

Different developers and different agents describe the same thing differently. Meaning-based retrieval is what makes a shared corpus usable.

It survives time

Terminology drifts as a codebase evolves. Embeddings degrade far more gracefully than exact-match rules.

// Commonly confused with

Commonly confused with.

Three neighbouring ideas.

Keyword searchMatches the words themselves. Fast and exact, and useless when the query and the entry describe one thing in two vocabularies.
Hybrid searchCombines both, using keyword matching for precise identifiers and semantic matching for everything else. Useful when exact tokens like a function name matter.
RAGA pattern that uses retrieval to inform generation. Semantic search is often the retrieval step inside it, not a synonym for the whole pattern.
// FAQ

Semantic search questions

Why not just use keyword search?

Because the query and the stored memory are usually written by different agents at different times in different words. Exact matching fails precisely when a corpus has become large enough to be valuable.

Does semantic search return wrong results?

It returns the nearest matches, which is not the same as correct. That is why ranking beyond raw similarity matters, and why stale entries need to expire rather than linger.

What happens as the corpus grows?

Without maintenance, quality drops as old entries compete with current ones. Virex Memory refreshes relevance scoring on a schedule and expires stale memories automatically.

Can I search memories myself?

Yes. The portal has a browse and search surface, so the corpus is not only reachable by agents.

Does search cost extra?

No. A seat is $15 per seat per month, with no per-query or per-token component.

Retrieval that survives your vocabulary.

Virex Memory searches by meaning across everything your agents have written.