Ask for what you meant, not what you typed.
Memories are embedded as vectors and retrieved by meaning. An agent looking for prior context does not have to reproduce the phrasing somebody used a month ago, and ranking keeps up as the corpus ages.
// Retrieval by meaning, ranked by a score that is refreshed on a schedule.
A store you cannot search is a store you will not use.
Memory only pays off at the moment of retrieval. Everything else is filing.
Nobody remembers the original wording
The memory that answers today question was written weeks ago by a different agent in different words. Exact-match search asks you to guess those words.
Agents cannot browse for context
An agent mid-task needs the relevant memory returned, not a list to skim. Retrieval has to do the judging.
Old memories crowd out current ones
A corpus that treats every entry as equally current gets worse as it grows, which is how a memory layer quietly stops being trusted.
Embedded on the way in, ranked on the way out.
Two mechanisms, working together: meaning-based matching, and a score that keeps ranking honest over time.
Vectors, not strings
Each memory is embedded as a vector when it is written, so comparison happens in meaning space rather than character space.
Composite relevance scoring
A scheduled scoring process keeps ranking current, so search results reflect what actually matters over time rather than raw recency or raw similarity alone.
Lifecycle cleanup underneath
Stale memories expire automatically, which keeps the pool being searched worth searching.
What retrieval behaves like.
The properties that matter when an agent is deciding what to trust mid-task.
Semantic search questions
How is semantic search different from keyword search?
Keyword search matches the words you typed. Semantic search matches what you meant, because memories are stored as vectors and compared by meaning rather than by exact string.
Do my agents have to phrase queries carefully?
No. That is the point of retrieving by meaning. An agent asking about a retry policy can surface the memory that described the behaviour without reusing its wording.
What is composite relevance scoring?
A scheduled scoring process that keeps search ranking current as memories age, so results reflect what actually matters over time rather than treating every memory as equally current.
Does search get worse as the corpus grows?
Relevance scoring is refreshed on a schedule and lifecycle cleanup expires stale memories automatically, so growth does not require a manual curation effort to stay useful.
Who can search my organization memory?
Members of your organization holding a Virex Memory seat, and the agents acting on their behalf. Isolation between organizations is enforced at the database layer.
Make the next session start informed.
Subscribe, assign a seat, and let your agents retrieve what earlier ones worked out.