Agent Memory that works with your existing Search Stack
Session Abstract
An agent that forgets you between sessions breaks a core UX promise. For our Agent Studio, we built memory so that it works across classic search, vectors, or both, based on what you already have. I’ll share our journey: what worked, what surprised us, and why 65% is sometimes the best score you can get.
Session Description
We built our Agent Studio to orchestrate LLMs across providers, powering e-commerce agents, support copilots, and SaaS assistants. When we added memory, we had a constraint: our customers run with classic keyword search, neural search, or both. We want to make the most of the latest tech, but we didn’t want to lock anyone out.
This brought practical challenges. How do you make the most of classic search for agent memory? What can tie-breaking ranking, numeric filters, or optional word boosting offer? How can LLM-generated freeform tags improve long-tail retrieval when boosted in queries?
We’ve seen builders spend years optimizing their search. Users have too: they know how to “speak Google.” We wanted to teach your agents how to tap into both, learning to query whatever backend exists rather than demanding vector-only infrastructure.
I’ll walk through our approach to hybrid memory annotations, conversation compression into multi-query payloads, a synthetic data pipeline for multilingual keyword extraction across 35 languages, and benchmarking the resulting memory system on LongMemEval.
You’ll leave with practical patterns for building new capabilities on existing search engines.
