DEV Community

Felipe 0liveira
Felipe 0liveira

Posted on AI-assisted

Agent memory is eating GraphRAG's lunch, and rerankers just got a lot cheaper

This digest covers RAG and GraphRAG developments from 2026-10-19 through 2026-11-09: new retrieval and reranking papers from arXiv and Hugging Face Daily Papers, agent-memory architecture writeups from Anthropic and LangChain, document-parsing and graph-tooling updates from LlamaIndex and Neo4j, and a production deep-dive from Latent Space.

πŸ”₯ Highlights

  1. The components of agent memory β€” Anthropic's own layered memory architecture, with data on when curation beats raw search. The components of agent memory
  2. Building Evidence-Grounded Agent Memory with OpenWiki and Graphs β€” a graph layer beats a wiki alone, without full GraphRAG overhead.
  3. Agentic GraphRAG and an explorer for your own stack β€” a runnable reference agent over a real knowledge graph, published today. Agentic GraphRAG and an explorer for your own stack
  4. ReΒ³Rank: Listwise Reranking with Reasoning-Conditioned Retrieval Heads β€” a 0.6B reranker beating 32B models, straight cost win.
  5. RAGCompass: Generalizable Retriever Routing via Reranking β€” a trainable router for picking the right retriever per query.

arXiv (cs.CL / cs.AI)

  • ArcRAG: Agentic Compositional Retrieval over Knowledge Graphs β€” 2026-11-04. Builds a query-specific view of a knowledge graph at inference time from the entities named in the question, then plans and repairs multi-hop traversals as Cypher queries. Practical value: a construction-agnostic retriever that works whether your graph came from a schema-guided or an open-extraction pipeline.
  • Typed-Centrality Grounding for Agentic GraphRAG: A Symbolic Approach for Agentic Reasoning β€” 2026-11-06. Computes centrality measures per entity type as a deterministic symbolic signal alongside text-similarity retrieval, targeting GraphRAG's habit of treating the graph as homogeneous. Reports more context-consistent answers on a real-world dataset β€” useful if your GraphRAG hallucinates relationships that pure text similarity misses.
  • QueryFlow: Multi-Hop Graph Retrieval with Adaptive Query Evolution β€” 2026-11-03. Updates the query representation during multi-hop traversal by folding query-conditioned graph summaries back into a shared query state. Reports double-digit gains on macro Recall@10 and LLM-Acc across four multi-hop QA datasets, addressing evidence that only becomes relevant after intermediate hops.
  • RAGCompass: Generalizable Retriever Routing via Reranking β€” 2026-11-03. Scores retrieved passages themselves to pick the best retriever per query, trained contrastively against the answers each passage set produces. Reports a 28% relative F1 gain over the best single retriever, holding up on unseen corpora/retrievers β€” a generalizable routing layer for teams running multiple indexes.
  • ReΒ³Rank: Listwise Reranking with Reasoning-Conditioned Retrieval Heads β€” 2026-11-02. A listwise reranker where a generator writes a reasoning trace first, then a scorer uses that trace to produce relevance scores. Beats 11 baselines across 9 benchmarks, with a 0.6B variant outperforming 32B rerankers β€” a serious latency/cost lever for production rerank stages.
  • DecisionRank: On-Policy Distillation for Decision-Model Rerankers β€” 2026-10-30. Distills a generative reranker's ranking knowledge into a non-autoregressive "decision model" via on-policy, item-space distillation. The 8B version matches its generative counterpart while running 16-25x faster.
  • Retrieving from Your Own Mistakes: On-Policy Training of Embedding Models for Multi-Hop Retrieval-Augmented Generation β€” 2026-10-29. Trains multi-hop retrievers on their own retrieval trajectories, including mistakes, labeling each visited state with the earliest missing gold passage. Improves evidence completeness on harder, out-of-distribution long-chain variants of HotpotQA and 2WikiMultiHopQA.
  • Reliability Stress Testing of RAG Retrieval under Query-Relevant Corpus Poisoning β€” 2026-11-03. Stress-tests the retrieval admission step against poisoned documents, showing query-tailored poison can fully compromise a dense retriever like BGE-small. Also evaluates overlap-based admission and MMR as mitigations β€” a concrete checklist for RAG exposed to untrusted corpora.
  • MEMOIR: A Training-Free Memory Lifecycle for Long-Horizon LLM Agents β€” 2026-11-05. Combines judge-gated consolidation with retrospective attribution that credits memories that actually supported past answers, plus an exploration bonus for under-used memories. Targets two real production pains directly β€” consolidation discarding needed evidence, and retrieval favoring stale memories β€” with no added LLM calls on the write path.
  • Same Simulation, Different Evidence: Where Retrieved Tool Output Costs LLM Agents Accuracy β€” 2026-10-19. Shows that when long tool outputs are saved to files and retrieved on demand, GPT-5.1/5.2 lose 7.9-11.3 accuracy points via "evidence substitution" (answering from prior output instead of the retrieved file), while Claude models show no detectable cost. Explicitly flagging the authoritative output recovers most of the gap β€” a model-specific warning for agentic retrieval over tool/file outputs.
  • Signed but Steerable: Measuring and Closing the Text-Vector Integrity Gap in Offline Retrieval-Augmented Generation β€” 2026-11-04. Shows that signed, precomputed embedding packs for offline/edge RAG only prove publisher identity, not that vectors match the text β€” cosine-tolerance checks and compressed-FAISS audits both fail to catch tampering. Proposes verifying quantization codes with a calibrated guard band, which caught attackers in testing β€” important if you ship or consume precomputed vector indexes.

Hugging Face Daily Papers

  • DeFT-RAG: Decomposing Multi-Hop Questions into Self-Contained Sub-Questions with Reusable Answers β€” 2026-10-27. Generates one self-contained multi-hop sub-question at a time, conditioned on prior answers, instead of planning all sub-questions upfront, caching answers across questions. Best accuracy among 13 baselines on MuSiQue, 2WikiMultiHopQA and HotpotQA while using fewer input tokens than other iterative methods.
  • Glance: The Generator Is Its Own Reranker β€” 2026-10-29. Collapses reranking and generation into one model by adding a small scoring head to the generator so all candidates are scored in a single shared-context pass. Beats a BGE reranker on SciFact/NFCorpus with gains holding up to 200 candidates β€” relevant if you want to cut a separate reranker service from your pipeline.
  • Cache Engineering for Agents β€” 2026-10-20. Proposes letting agents directly manage their active KV cache instead of rewriting context and re-prefilling, plus an "AKV" serving engine exposing a stateless API over stateful cache reuse. Reports better accuracy, lower cost and lower P95 latency than threshold-triggered summarization β€” directly actionable for long agentic/RAG sessions hitting KV-cache cost walls.
  • EvoMemState: Evolving Memory State for Personal Agents β€” 2026-11-02. Treats agent memory as evolving relational state β€” tracking validity, scope and relations between facts over time β€” plus a separate reasoning state deciding how to apply facts to the current request. Targets the common gap where fact-centric memory misses that new information updates or contradicts older facts.

Anthropic Engineering Blog

  • The components of agent memory β€” 2026-11-06. Describes a layered memory architecture β€” searchable raw transcripts, memory notes, curated wikis, and a small always-loaded context layer β€” plus "dreaming," an offline process that corrects notes, prunes wikis, and promotes key facts upward. Benchmarks show a wiki-plus-search setup beats search alone on multi-document questions, but an oversized wiki can become less useful than the raw source β€” a concrete data point for choosing between pure vector search and curated-memory layers.

LangChain Blog

  • Context Engineering Starts with Retrieval β€” 2026-10-23. Argues agents fail mainly from retrieval problems β€” wasteful searches, oversized results, stale or irrelevant content β€” and lays out fixes used in LangChain's Deep Agents harness: filesystem-style search, semantic retrieval, paged reading, offloading results to files, and filtering subagents. A practical checklist for moving from naive two-step RAG to agentic retrieval without blowing the context budget.
  • Building Evidence-Grounded Agent Memory with OpenWiki and Graphs β€” 2026-10-28. Tests OpenWiki, an open-source wiki compiler, as an agent memory layer on a synthetic insurance-document corpus, then adds a lightweight evidence graph turning wiki citations into typed edges an agent traverses via locate/expand tools. The graph layer improved accuracy, citation consistency, and cost-per-correct-answer versus the wiki alone β€” a cheap, extraction-light alternative to full GraphRAG worth evaluating before building a heavier knowledge-graph index.
  • How Numan built an eval-first AI health coach with LangGraph and LangSmith β€” 2026-10-19. A case study on why single-turn fixtures and scripted multi-turn tests gave misleading signal, and how the team moved to per-category LLM-judge scoring, simulated patient conversations, and clinician-owned rubrics gating releases. Directly actionable for teams whose RAG/agent evals rely on static test sets that don't catch conversational drift or safety regressions.

LlamaIndex Blog

  • OCRia-1: 4B OCR Model for Document Parsing β€” 2026-10-22. Introduces OCRia-1, a 4B-parameter vision-language OCR/parsing model trained on 100K+ pages, reporting state-of-the-art results for sub-35B models on ParseBench and strong scores on table/chart extraction. Upstream of RAG: better table/chart/formula fidelity in parsing directly affects chunk quality and retrievability of structured data.
  • Introducing the Document Parsing Protocol β€” 2026-10-28. Proposes an open spec (working draft 0.1, ~140 rules) defining how parsers should structure tables, forms, charts, and diagrams, grown out of labeling guidelines for the ParseBench benchmark. Gives document-ingestion teams a concrete target format to evaluate parser output against before it reaches chunking and embedding.
  • The Inference Stack Behind OCRia β€” 2026-11-04. Details how LlamaIndex got roughly 10x throughput (285 to 2,800 tok/s) on a single B200 via mixed-precision/activation-aware quantization, speculative decoding, and kernel fusion. Practical for anyone weighing self-hosted document-parsing cost and throughput against hosted OCR APIs.

Neo4j Blog

  • Agentic GraphRAG: How an AI agent uses a knowledge graph β€” 2026-10-19. Explains agentic GraphRAG as a reason-act-observe loop where the agent chooses graph entry points, query tools, and traversal depth at runtime, and can extend the graph as new information arrives via a reusable context graph. A good conceptual reference for deciding when agentic traversal is worth the added complexity over static graph retrieval.
  • Neo4j Context Engine: Rapidly build dependable enterprise agents β€” 2026-10-20. A private-preview product built on a six-layer graph-based business ontology connecting business meaning, processes, and data sources, with components for ontology management, GraphRAG-based governed retrieval across systems, and storage of agent decisions. A reference architecture for enterprise GraphRAG with data governance built in.
  • The Neo4j Ontology Graph Specification: A pragmatic, open, AI-first approach β€” 2026-11-05. An Apache-2.0 spec (YAML/JSON, validated by JSON Schema) for defining nodes, relationships, typed properties, cardinality, and links to external standard terms, with SDKs for JVM, JS, and Go. Useful for standardizing graph construction and ontology design across tools in a GraphRAG indexing pipeline instead of hand-rolling schemas.
  • Agentic GraphRAG and an explorer for your own stack β€” 2026-11-09. Introduces graph-explorer-kit, a Python library on top of neo4j-graphrag that runs an agent over a knowledge graph built by Neo4j's Document Intelligence feature, exposed via an HTTP API and a browser-based explorer UI, demonstrated on 240 documents. A concrete, runnable reference implementation teams can adapt directly.

Simon Willison

Nothing with real technical substance on retrieval, indexing, reranking, embeddings, or agent memory this window β€” the period's posts were mostly tool-release notes (llm version bumps, datasette-agent).

Latent Space

  • Building AI for Lawyers - Jacob Lauritzen, CTO of Legora β€” 2026-11-02. Legora's CTO frames legal research as a fan-out retrieval problem solved via sub-agent delegation rather than single-shot search, describing a move from MapReduce-style chunking toward large-context processing for million-token contracts (chunking loses nuance across chunk boundaries) and the use of turbopuffer's multivector support for late-interaction retrieval. A real-world data point on when to abandon chunking for long-context ingestion, and a vote of confidence for late-interaction retrieval as production-ready.
  • How Anthropic ships Claude with Claude: Separating the Brain from the Hands β€” 2026-10-27. Covers Anthropic's internal "Managed Agents" and a multi-agent "Dreaming" pattern, where one agent reviews another's prior work and writes memories back into it, plus notes on preserving reasoning blocks across turns under context compaction. A working pattern for cross-session agent memory write-back and a heads-up on compaction tradeoffs when long agent traces must stay coherent.

Interconnects

Nothing relevant this window β€” the archive's recent entries covered open-model economics, benchmarking, and geopolitics, not retrieval or agent memory.

Through-line

Agent memory is this window's center of gravity, and it's quietly absorbing work that used to be GraphRAG's job. Anthropic's layered memory design, LangChain's graph-backed OpenWiki experiment, and Latent Space's look at Anthropic's internal "Dreaming" pattern all converge on the same idea: a curated, periodically-rewritten memory layer with light graph structure beats either raw search or a heavyweight knowledge graph built up front. On the retrieval-mechanics side, the arXiv crop shows reranking getting much cheaper (ReΒ³Rank, DecisionRank) and routing/evolution logic getting smarter (RAGCompass, QueryFlow) rather than chasing bigger single models β€” the cost curve for "good enough" retrieval keeps bending down.

What's catching your eye in agent memory versus GraphRAG right now β€” are you consolidating toward lighter memory layers too? Drop a comment.

Top comments (0)