Importance scoring
Also called: relevance scoring, memory decay, recency weighting
What is importance scoring in agent memory?
Importance scoring ranks stored memories by how much they still matter, and refreshes those ranks on a schedule. A note about a module that was deleted should sink; a constraint the team keeps hitting should not. Scoring at write time alone cannot know either.
Why the score cannot be set once.
Importance is a property of now, and the moment of writing is a bad time to judge it.
The writer cannot know
An agent recording a conclusion mid-task has no idea whether this turns out to be a recurring constraint or a one-off. Asking it to assign a permanent importance is asking it to predict.
Relevance moves as the codebase moves
Code gets deleted, decisions get reversed, dependencies get replaced. A memory that was central can become misleading without its text changing at all.
Repeated use is evidence
A memory that keeps getting retrieved and keeps proving useful is demonstrating its own importance. That signal only exists after the fact.
Why it is the difference between a corpus and an archive.
Storage is easy. Deciding what should surface is the product.
Retrieval has to choose
Context space is finite, so something must decide which of many relevant memories actually goes in.
Stale beats missing, but not by much
A confidently wrong memory about deleted code is worse than no memory, which is why decay matters.
The corpus should get sharper
Use is signal. A store that learns from its own retrieval history improves rather than merely enlarging.
Commonly confused with.
Three neighbours.
Importance scoring questions
What is importance scoring in agent memory?
Ranking stored memories by how much they still matter, refreshed on a schedule, so retrieval reflects current relevance rather than a judgement made at the moment each entry was written.
Why not score once at write time?
Because the writer cannot know. An agent recording a conclusion has no way to tell whether it is a recurring constraint or a one-off, and relevance moves as the codebase moves.
Is this just recency weighting?
No. Recency is one input. The oldest memory in a corpus can easily be the most important thing in it.
What happens to stale memories?
They sink. A confident memory about deleted code is worse than no memory, so decay is part of keeping the corpus trustworthy.
How is it different from search relevance?
Search relevance is similarity to one query. Importance is a property of the memory itself, independent of what was asked.
Memory that stays relevant.
Virex Memory re-scores importance on a schedule instead of trusting a guess made at write time.