History — 2026 week 37 (Sep 7 – Sep 13, 2026)
Newest first.
2026-09-07 — Diversity Measured by Aspect, Not Distance (7 new, 3 updated)
Ten news articles about one event can sit far apart as documents while repeating the same three facts, which is why diversifying over whole-document embeddings — what MMR does — leaves the redundancy in place. NEWSCOPE moves the unit of diversity down to the sentence, clustering candidate sentences into event aspects and picking paragraphs greedily by how many uncovered clusters each brings. Its two coverage-based metrics separate systems that APD cannot: on LocalNews the reranker lifts aspect coverage from 53 to 75 while sitting level with MMR on average pairwise distance. EMNLP 2025; benchmarks ship with the code, unlicensed, and every figure is the authors’ own.
Articles — Uncovering the Bigger Picture - Comprehensive Event Understanding via Diverse News Retrieval People — Yixuan Tang · Yuanyuan Shi · Yiqun Sun · Anthony K.H. Tung Datasets — LocalNews (Google News full-coverage events) · DSGlobal (adapted from DiverseSumm) Updated — APD · Diversity Metrics · Search Result Diversity
2026-09-07 — Clickstream Judgments Without the Engine (1 new, 5 updated)
Deriving relevance judgments from click data has meant running OpenSearch, because Search Relevance Workbench was the only evaluation tool with UBI capture built in. Releval breaks that pairing: engine-agnostic like Quepid, pointing at Elasticsearch, Solr, Vespa or any HTTP endpoint, but collecting UBI queries and events itself into ClickHouse, with LLM judges and an MCP server alongside. Relevance Evaluation Tools Compared now weighs it against the established options. Two things set it apart from everything else there: it is proprietary and EULA-gated, and version 1.0.0 shipped in July 2026. Judgment uploads follow the TREC qrels shape.
Tools — Releval Updated — Relevance Evaluation Tools Compared · User Behavior Insights · Quepid · Rated Ranking Evaluator · Search Relevance Workbench
2026-09-07 — Preferences Extracted From Conversation (6 new, 5 updated)
DICE takes the position that an agent’s memory belongs in the application’s own domain model rather than a new vector store, and that code — not the model — decides what enters the context window. Conversation turns are extracted into natural-language propositions scored for confidence, importance and decay, reconciled against what is already stored, and the top few by importance × confidence are prepended to the next prompt. The lineage is the General User Models paper, supplying observation-derived propositions, confidence, revision and staleness. Embabel DICE is the Kotlin reference implementation; django-dice is Andrew Kornilov’s Python port, and Jettro Coenradie shows the wiring end to end on Embabel, whose extraction prompt forbids inference beyond the evidence.
Articles — Agents That Extract and Use Preferences from Conversations Concepts — DICE People — Jettro Coenradie Tools — Embabel · Embabel DICE · django-dice Updated — Agentic Memory · Context Engineering · Agentic Memory for Search Personalization · E-commerce Search · Andrew Kornilov
2026-09-07 — Building Personalized Search on Memory (1 new, 7 updated)
Holding a user model as language instead of as learned features changes where the engineering goes, and Agentic Memory for Search Personalization now works that through end to end: which behavioural events are worth keeping, why preference inference belongs offline and query-time application online, and the three places preference can enter — rewrite, rerank, filters — in increasing order of harm when wrong. It sets the OpenSearch and Elasticsearch designs side by side on isolation, where one partitions and the other enforces, and is candid that per-user rewriting leaves evaluation, cold start and attribution unresolved.
Topics — Agentic Memory for Search Personalization Updated — Agentic Memory · Personalization · Personalization in Search · Conversational and Agentic Search · Agentic Memory - OpenSearch Docs · AI Agent Memory - Creating Smart Agents with Elasticsearch Managed Memory · Hajer Bouafif - Personalize Search Results with OpenSearch Agentic Memory
2026-09-07 — What OpenSearch’s Memory Types Actually Hold (1 new, 4 updated)
The reference documentation for Agentic Memory organises everything into a memory container — one bundle of embedding model, LLM, extraction strategies and namespaces per use case — and its four memory types divide by stage of processing, not by lifetime: raw content lands in working and is promoted to long-term only when the infer flag is set, which is off by default. Agentic Memory - OpenSearch Docs records the API surface, the SEMANTIC/USER_PREFERENCE/SUMMARY strategies, and an unusually direct notice that namespaces partition memories but access control remains the operator’s job.
Corrections
- Agentic Memory described
sessionsmemory as holding the current conversation; it holds session metadata, while recent messages and execution traces areworkingmemory. - Agentic Memory described
historyas an audit trail of the agent’s actions and reasoning; it logs memory operations — add, update, delete. - The OpenSearch 3.3 date, previously carried as Hajer Bouafif’s claim, is confirmed by the documentation; the 3.5 context-management date is not covered there and stays attributed to her talk.
Articles — Agentic Memory - OpenSearch Docs Updated — Agentic Memory · OpenSearch · Personalization · Multi-Tenancy in Search
2026-09-07 — Three Agents to Tune Query DSL (4 new, 5 updated)
Relevance tuning normally needs someone fluent in the query DSL. The OpenSearch Relevance Agent splits that job across three agents — one reading UBI data to locate relevance gaps, one turning findings into tuning strategies, one testing them against offline evaluation sets — and drives the whole thing from a chat box in Dashboards, claiming diagnosis in hours rather than days. Two design choices carry the weight: metric computation is held outside the LLM in deterministic tools so numbers are measured, and every agent reaches the engine only through the MCP server. Experimental in OpenSearch 3.6, confined to query-DSL edits. From Bobby Mohammed and Daniel Wrigley.
Articles — Introducing OpenSearch Relevance Agent Concepts — Model Context Protocol People — Bobby Mohammed Tools — OpenSearch Relevance Agent Updated — OpenSearch · User Behavior Insights · Search Relevance Workbench · Daniel Wrigley · Agentic Search
2026-09-07 — Agent Memory, and Who Is Allowed to Read It (3 new, 6 updated)
An agent’s memory is not one thing. Procedural memory — when to store, when to retrieve, how to use tools — is harness code and never reaches the index; episodic and semantic memory do. AI Agent Memory - Creating Smart Agents with Elasticsearch Managed Memory builds the split on Elasticsearch and then makes isolation the engine’s problem: a role descriptor pins a filter onto the memory’s type field, so credentials decide which memories exist at all and the agent issues the same query either way, over a retriever fusing ELSER and keyword with RRF. Agentic Memory now carries this taxonomy alongside the lifetime-based one. From Gustavo Llermaly and Jeffrey Rengifo.
Articles — AI Agent Memory - Creating Smart Agents with Elasticsearch Managed Memory People — Gustavo Llermaly · Jeffrey Rengifo Updated — Agentic Memory · Elasticsearch · ELSER · Elastic · Reciprocal Rank Fusion · Agentic Search
2026-09-07 — Memory Instead of Features for Personalization (4 new, 11 updated)
Three shoppers typing “black shoes” may want heels, sneakers or work boots, and no amount of ranking work resolves which of them is asking. Agentic Memory holds that answer as language rather than as a coefficient: preferences inferred offline by an LLM from UBI clickstream and written to long-term memory as facts, then read at query time and applied through Query Rewriting, where “backpack” becomes “backpack leather neutral professional”. A returned item whose feedback says “too flashy for my work environment” teaches a feature-based ranker to decay the colour orange, losing both the reason and the alternative. From Hajer Bouafif’s Berlin Buzzwords 2026 talk; she dates OpenSearch’s four memory types to 3.3.
Videos — Hajer Bouafif - Personalize Search Results with OpenSearch Agentic Memory Concepts — Agentic Memory · Query Rewriting People — Hajer Bouafif Updated — OpenSearch · User Behavior Insights · Personalization · Agentic Search · Query Understanding · Query Expansion · Query Relaxation · Personalization in Search · Conversational and Agentic Search · Berlin Buzzwords · Amazon Web Services