Hajer Bouafif

OpenSearch Solutions Architect at Amazon Web Services. Works on Personalization and Agentic Memory in OpenSearch, and presented the memory-based approach to search personalization at Berlin Buzzwords 2026.

Her argument is that personalization has been mis-framed as a ranking-and-feature problem. The question a personalized system should answer is not how do we better rank these results but who is running this query, and why now — and the durable answer to that lives in memory rather than in a retrained feature space. She positions memory as the agent’s “source of intelligence”, complementary to grounding: the catalog tells the agent what exists, the memory tells it what this particular user means.

A second recurring theme is operational discipline around agents. Her production guidance keeps LLM reasoning off the query path entirely — preferences inferred offline, profiles cached, retrieval capped at the top 5–10 facts, a small language model online, and a conventional lexical/k-NN/hybrid fallback always standing by for when the agent is slow or misbehaves.

Talks in this vault

Topics