retrieval
We Added Jev as a Reranker. Here's What We Learned.
Six things we learned adding Jev, a model that returns typed decisions instead of text, as a reranker in Hindsight 0.10.1. Including two designs that failed and the numbers that killed them.
Stop Growing Your Always-On Context
The default fix for an agent that forgets is to put more in the prompt. It works until it doesn't, and the way it fails is quiet: unconditional context that costs every turn and gets less relevant as it grows.
Cross-Encoder Reranking: The Last Stage of Agent Memory Recall
Retrieval hands the reranker a few hundred plausible memories. What happens next decides what your agent actually sees. Inside Hindsight's cross-encoder stage: score normalization, multiplicative boosts, candidate budgets, and how to make recall fail open.
Knowledge Graphs vs. Vector Search for Agent Memory
Should your agent's memory be a vector store or a knowledge graph? They solve different halves of the problem. Here is where each one wins, where each breaks, and why agent memory needs both.
Entity Resolution in Agent Memory: One Person, Many Names
How agent memory decides that Sarah, Sarah Chen, and she are one person, and why Hindsight resolves entities with a co-occurrence graph, not embeddings.
How We Built a 4-Way Hybrid Search System That Actually Runs in Parallel
Sequential async queries were killing our retrieval latency. Here's how we built a true 4-way parallel hybrid search system with asyncio and RRF fusion — then evolved it further with connection sharing, cross-encoder reranking, and multiplicative boost scoring.
How We Built Time-Aware Spreading Activation for Memory Graphs
Hindsight — Stories, Not Rows