The Problem
Embedding-based search is fast. It finds memories that are semantically similar to your query. But “similar” isn't always “right.”
When your AI CMO asks “what did the client say about their brand voice?” — vector search might return 20 memories about brand voice in general. The re-ranker finds the three that are about this client, from this onboarding call, with their specific preferences.
That precision is the difference between a generic response and one that sounds like it came from a team member who was in the room.
The Pipeline
The embedding engine casts a wide net — pulling the top 50-100 candidate memories from vector search. Fast, but not precise enough for high-stakes decisions.
Every candidate is scored pairwise against your exact query using a cross-encoder model. This catches semantic nuances that embedding similarity misses — the difference between "we tried this" and "we tried this and it failed."
The top-ranked memories are injected into the agent's working context with full provenance — who stored it, when, and why. Your agent doesn't just remember. It remembers with confidence.
Capabilities
A CMO retrieving brand memories gets different rankings than a CTO retrieving architecture decisions. The re-ranker understands agent roles and surfaces what's relevant to each.
Re-ranking happens in under 100 milliseconds, even across thousands of memories. Fast enough for real-time agent workflows without blocking execution.
Identity, expertise, and experience memories are weighted differently based on the query type. Factual questions pull from expertise. Tone questions pull from identity. The system knows the difference.
Not just what's most similar — what's most useful right now. Recent corrections outrank old habits. Fresh decisions outrank stale preferences. Time-aware ranking prevents agents from acting on outdated information.
Comparison
You don't buy re-ranking separately. It's part of the memory infrastructure that powers your AI department. Book a call to see it in action.