Daily Brief — Jul 13, 2026: Agent Memory

🛰 AI Brief — Jul 13, 2026

How to read prio and sources

prio N is the radar’s practical-relevance score for this item (higher runs first; items at or below the noise threshold are filtered out as noise). Under each signal: Concepts / Entities are graph links; Source / N sources list every outbound link for that story.

🥇 Clinical RAG paper finds entity-attribution failures that standard evals miss · prio 12

For builders working on RAG systems, this is a concrete warning that faithfulness and citation checks are not enough if the evidence can be attached to the wrong entity. The paper also points to a specific missing evaluation layer, entity-attribution verification, that the authors say no existing framework implements. Concepts: RAG RAG Evaluation Entities: arXiv Source: arxiv.org

🥈 Shared Selective Persistent Memory for Agentic LLM Systems · prio 12

Builders working on agentic workflows because it focuses on what to keep across sessions, what to drop, and how to share workspace state safely. The paper also gives concrete evidence that naive full-history persistence can hurt completion, which is a useful warning for anyone designing memory-heavy agent systems. Concepts: Agent Memory Context Engineering Agents Tool Use Entities: arXiv 2 sources: arxiv.org, arxiv.org

🥉 Claude Code’s effort setting is about work done, not just thinking time · prio 12

For builders using Claude Code, the key distinction is practical: model choice changes the underlying capability, while effort changes how much of the available context and verification work the system is willing to spend on the task. The article gives a concrete mental model for why a request can behave differently even when the model name stays the same. Concepts: Context Engineering Entities: Fable Sonnet Source: habr.com

4️⃣ MemDocAgent uses shared memory and dependency-aware traversal for repository-level code documentation · prio 11

The paper is directly about a workflow problem that coding-agent builders face: producing consistent documentation over a whole repository without redundant retrieval or conflicting outputs. Its main technical takeaway is the combination of traversal order plus shared memory for tracking prior work traces, which is relevant to anyone building agentic code workflows. Concepts: Agents Agent Memory Code Agents Context Engineering Tool Use Source: arxiv.org

5️⃣ LongMedBench: Benchmarking Medical Agents for Long-Horizon Clinical Decision-Making · prio 11

Builders working on agents and retrieval systems because it tests long-horizon behavior, temporal reasoning, and memory over real EHR-derived interaction streams. It also gives the community a concrete evaluation frame for when retrieval or agent memory helps, and where immediate context still dominates decision-making. Concepts: RAG Agent Memory LLM Evals Long Context Source: arxiv.org

Knowledge Gaps

Topics the AI stream keeps raising that the knowledge base hasn’t sufficiently covered yet — candidates for what to learn next. Agent Memory · Reranking · RAG · Embeddings · Context Engineering

🚀 Models & Releases (1)

• prio 7 Soofi S 30B-A3B: open-source German-English foundation model with hybrid MoE design Concepts: Long Context Open Source LLMs LLM Evals Entities: Deutsche Telekom German Industrial AI Cloud Soofi S 30B-A3B Olmo-3-32B Source: arxiv.org

🧪 Research Papers (73)

• prio 11 Eluna: An agentic LLM system for warehouse SOP execution Concepts: Agents Tool Use Context Engineering Source: arxiv.org
• prio 11 SAGEAgent proposes a self-evolving clinical agent for cost-aware diagnostic acquisition Concepts: Agents Agent Memory Tool Use Source: arxiv.org
• prio 11 AutoMem trains AI agents to manage memory as a skill Concepts: Agent Memory Entities: Stanford University Source: qudata.com
• prio 10 WILDTRACE benchmarks long-context reasoning over naturally dispersed evidence Concepts: LLM Evals Long Context 6 sources: arxiv.org, arxiv.org, arxiv.org, arxiv.org, arxiv.org, twitter.com
• prio 10 Deco-G separates task solving from output formatting in LLM decoding Concepts: LLM Evals Source: arxiv.org
• prio 10 REAL proposes failure-aware KV cache eviction for long-context models Concepts: Long Context LLM Evals 2 sources: arxiv.org, arxiv.org
• prio 10 RAG vs long-context prompting for EHR clinical reasoning Concepts: RAG Long Context Entities: GPT-5.4-mini Mistral Medium 3 DeepSeek-V3.1 Source: arxiv.org
• prio 10 InfoNCE paper links negative samples to similarity-search generalization Concepts: Embeddings Source: arxiv.org
• prio 10 AgentKGV: Agentic LLM-RAG for Knowledge Graph Fact Verification with Two-Stage Training Concepts: RAG Agents Tool Use Source: arxiv.org
• prio 10 GRACE proposes graph-based verification for long-horizon agent instruction updates Concepts: Context Engineering Agents Entities: Google Gemini 2.5 Flash Gemini-3.1-Pro 2 sources: arxiv.org, habr.com
• prio 10 Self-Guided Test-Time Training for Long-Context LLMs Concepts: Long Context Entities: Qwen3-4B-Thinking-2507 Llama 3.1 8B Instruct Source: arxiv.org
• prio 10 The Patchwork Problem in LLM-Generated Code Concepts: Code Agents LLM Evals Source: arxiv.org
• prio 10 Evaluating Energy, Performance, and Accuracy Trade-offs Across vLLM Configurations Entities: vLLM Source: arxiv.org
• prio 10 Anatomy of CLI Coding Agent Failure Trajectories Concepts: Code Agents LLM Evals Entities: DAIR.AI Source: twitter.com
• prio 9 Small hyperbolic language models are used to study creativity, honesty, and designed forgetting Concepts: LLM Evals Agent Memory Source: arxiv.org
• prio 9 Geopolitical endorsement changes LLM policy judgments Concepts: LLM Evals Entities: OpenAI Anthropic Google DeepSeek Source: arxiv.org
• prio 9 MedRealMM benchmarks multimodal online medical consultation on real patient-doctor cases Concepts: LLM Evals Entities: Hugging Face Source: arxiv.org
• prio 9 Forget Narrowly, Retain Broadly: Unlearning as an Asymmetric Generalization Problem Concepts: LLM Evals Source: arxiv.org
• prio 9 AutoWorldBuilder paper on multi-agent worldbuilding with context compression and iterative review Concepts: Agents Context Engineering LLM Evals Entities: GPT-OSS 120B DeepSeek-V3.2 Source: arxiv.org
• prio 9 Memory-managed long-context attention with bounded editable memory and sparse fallback Concepts: Long Context RAG Entities: LLaMA Qwen Source: arxiv.org
• prio 9 SCATE learns automated supervision for coding-agent test generation Concepts: Code Agents Agents Source: arxiv.org
• prio 9 QuantCode-Bench evaluates whether LLMs can generate executable trading strategies Concepts: LLM Evals Source: habr.com
• prio 8 Paper adds dialogue to PARTNR for embodied multi-agent coordination Concepts: Agents LLM Evals Source: arxiv.org
• prio 8 ArXiv paper proposes source-level recovery of stripped binaries using retrieval and LLM reasoning Concepts: RAG Reranking LLM Evals Entities: GitHub arXiv Source: arxiv.org
• prio 8 A Personalized Computational Framework for Assessing the Sufficiency of Partially Observed Data in Healthcare AI Models Source: arxiv.org
• prio 8 Git-Assistant uses planning to support non-trivial Git operations Concepts: Agents Tool Use Code Agents Context Engineering Source: arxiv.org
• prio 8 Hierarchical Chain-of-Thought improves reasoning accuracy and trace length in benchmark tests Concepts: LLM Evals Source: arxiv.org
• prio 8 PhysAssistBench evaluates LLMs for interactive doctor-patient-EHR assistance Concepts: LLM Evals Agents Tool Use Source: arxiv.org
• prio 8 Correlation-Aware Contextual Bandits with Surrogate Rewards for LLM Routing Source: arxiv.org
• prio 8 Director proposes online proactive expert placement for distributed MoE serving Entities: arXiv Mistral DeepSeek Qwen Source: arxiv.org
• prio 8 TheBioCollection: a 52.6B-token biology pretraining corpus with a matched evaluation suite Concepts: LLM Evals Entities: arXiv Gravity-16B-A3B Source: arxiv.org
• prio 8 SLIDERS automates evidence synthesis for systematic reviews Concepts: RAG Agents LLM Evals Source: arxiv.org
• prio 8 KV-PRM proposes KV-cache-based process reward modeling for multi-agent test-time scaling Concepts: Agents LLM Evals Source: arxiv.org
• prio 8 LDT-Coord Uses a Digital Twin to Cut Coordination Traffic in Heterogeneous LLM Agent Teams Concepts: Agents Tool Use Source: arxiv.org
• prio 8 Present but Rescaled: Chat-to-Agent Transfer of Additive Activation Steering Concepts: Agents Tool Use Entities: Qwen2.5-7B Gemma-2-9B Yi-1.5-9B Source: arxiv.org
• prio 8 Beyond Fixed Representations: The Vocabulary and Verifier Gaps in Open-Ended AI Concepts: LLM Evals Source: arxiv.org
• prio 8 Anthropic studies how Claude’s expressed values vary across models and languages Concepts: LLM Evals Entities: Anthropic Claude.ai Claude Sonnet 4.6 Source: anthropic.com
• prio 8 Anthropic finds Claude’s tone changes by language, with Russian scoring as the strictest Concepts: LLM Evals Entities: Anthropic Claude Sonnet 4.6 Opus 4.6 Source: habr.com
• prio 7 Paper finds that forcing formal explanations hurts MLLM classification accuracy Concepts: LLM Evals Source: arxiv.org
• prio 7 Multi-Agent Troubleshooting for Telecom Networks Uses a Fine-Tuned SLM for Remediation Planning Concepts: Agents Tool Use Source: arxiv.org
• prio 7 Test-time scaling on small multilingual VLMs shows parsing and decoding budget matter more than search Concepts: LLM Evals Entities: Qwen2.5-VL-7B-Instruct Qwen3.5 4B Source: arxiv.org
• prio 7 Sensitivity-Aware Thresholding and Token Routing for LLM Activation Sparsification Source: arxiv.org
• prio 7 Study Maps How Software Engineering Skills Are Being Packaged for Agents Concepts: Agents Tool Use Source: arxiv.org
• prio 7 Sticky Routing for MoE Training Aimed at More Memory-Efficient Inference Source: arxiv.org
• prio 7 Super-Tuning proposes sparse PEFT support selection from pruning-style saliency Entities: Meta Llama-3.2-1B Meta-Llama-3-8B Source: arxiv.org
• prio 7 Mach-Mind-4-Flash technical report on a 35B MoE agentic model Concepts: Agents LLM Evals Entities: Mach-Mind-4-Flash 2 sources: arxiv.org, arxiv.org
• prio 7 A GPU inference method for moderately sparse LLM weight matrices Source: arxiv.org
• prio 7 Diversify2Verify studies how implementation structure affects automated program verification Concepts: LLM Evals Source: arxiv.org
• prio 7 OpenProver adds an open-source, interactive Lean 4 theorem-proving system Concepts: Agents LLM Evals Source: arxiv.org
• prio 7 REFORGE benchmarks LLM reverse-engineering performance on decompiled binary function naming Concepts: LLM Evals Source: arxiv.org
• prio 7 Agora: Auction-Based Task Allocation for LLM Agents Concepts: Agents Tool Use LLM Evals Source: arxiv.org
• prio 7 AI systems as semantic abstractions and authority-bound representations Concepts: Agents Tool Use LLM Evals Source: arxiv.org
• prio 7 NL-PAC: Specification Ambiguity and Certified Minimax Risk Floors in LLM-Mediated Supervision Concepts: LLM Evals Entities: arXiv Qwen-2.5 3B Source: arxiv.org
• prio 6 Attributing vision-language model errors before decoding Source: arxiv.org
• prio 6 Beyond Black-Box Obfuscation: Mechanistic Analysis and Defense of White-Box Monitors Concepts: LLM Evals Entities: Anthropic arXiv Source: arxiv.org
• prio 6 Mechanistic explanation for why memorized facts can still fail in LLM finetuning Entities: alphaXiv Source: twitter.com
• prio 6 PRecG: Legal Precedent Retrieval with Graph Neural Networks and Rhetorical Role Segmentation Concepts: RAG Embeddings Source: arxiv.org
• prio 6 OmniMapBench benchmarks visual-centric reasoning on map documents Concepts: LLM Evals Source: arxiv.org
• prio 6 An Emergent Mirage: Is Emergent Misalignment and Realignment Indeed a Robust Phenomenon? Concepts: LLM Evals Source: arxiv.org
• prio 6 Evolutionary intelligence framework for cumulative scientific discovery Entities: arXiv alphaXiv Connected Papers Litmaps Source: arxiv.org
• prio 6 Pattern-Aware Graph Neural Networks for Handling Missing Data Source: arxiv.org
• prio 6 Activation-Guided Suffix Attacks Probe Refusal Representations in Language Models Concepts: LLM Evals Source: arxiv.org
• prio 6 Toward Auditable AI Scientists: A Hypothesis Evolution Protocol for LLM Agents Concepts: Agents Tool Use Source: arxiv.org
• prio 6 Accelerating LLM inference with self-supervised early exits Entities: Pythia Source: arxiv.org
• prio 6 Paper on using GPT-4o and RAG to generate investor briefs from company, macro, and SEC data Concepts: RAG Entities: OpenAI U.S. Securities and Exchange Commission EDGAR GPT-4o Source: arxiv.org
• prio 6 The LM head may be a gradient bottleneck in language model training Source: arxiv.org
• prio 6 Signed Symmetric Quantization for Few-Bit Integers Entities: AMD Qwen3 Qwen3.5 Llama3 Source: arxiv.org
• prio 6 ProofCouncil: an LLM agent for open mathematical problems Concepts: Agents LLM Evals Source: arxiv.org
• prio 6 EvoLP proposes a self-evolving latency predictor for edge model compression Source: arxiv.org
• prio 6 Multimodal Reward Hacking in Reinforcement Learning Concepts: LLM Evals Agents Source: arxiv.org
• prio 6 STEEL brings FlashAttention-style inference to AMD XDNA NPUs Entities: AMD Source: arxiv.org
• prio 6 TrustX ARC proposes a three-tier risk framework for agentic AI systems Concepts: Agents Code Agents Entities: TrustX arXiv Source: arxiv.org
• prio 6 Reinforcement Learning from Hindsight for Vision-Language-Action Models Entities: arXiv Franka Source: arxiv.org

🛠 Tools & Frameworks (12)

• prio 9 AnySearch tops Product Hunt as an agent-focused search product Concepts: Agents Tool Use MCP RAG Entities: AnySearch Product Hunt Frames FreshQA 9 sources: qbitai.com, arxiv.org, habr.com, github.com, bleepingcomputer.com, qbitai.com, qbitai.com, habr.com, qbitai.com
• prio 9 Clawk puts coding agents in a disposable Linux VM Concepts: Agents Tool Use Code Agents Context Engineering Source: github.com
• prio 9 Apple’s SpeechAnalyzer benchmarked against Whisper on-device Concepts: LLM Evals Entities: Apple OpenAI Inscribe Whisper Source: get-inscribe.com
• prio 8 Report says Grok Build CLI uploaded full Git repositories to Google Cloud storage Entities: xAI Google Cloud Storage Google Cloud International Cyber Digest Source: internationalcyberdigest.com
• prio 8 AI-powered Big Mouth Billy Bass built with Strands Agents and Amazon Nova 2 Sonic Concepts: Agents Tool Use Entities: Amazon AWS Anthropic Amazon Nova 2 Sonic Source: github.com
• prio 7 QQ: A language metadata toolkit for multilingual NLP Entities: Hugging Face arXiv Source: arxiv.org
• prio 7 HyperAgent skill builds an LLM verification wiki from papers Concepts: Agents Tool Use Context Engineering Entities: DAIR.AI HyperAgent Source: twitter.com
• prio 6 shot-scraper 1.11 adds JS file loading, timeout options, and slower server startup handling Source: simonwillison.net
• prio 6 colibrì ranks #6 in MoE with a pure-C GLM-5.2 runtime Entities: GLM-5.2 Source: github.com
• prio 6 DOM-docx converts semantic HTML fragments into editable Word documents Entities: Hacker News LibreOffice Playwright Chromium Source: github.com
• prio 6 ChatGPT returns to WhatsApp in the EEA and expands to Kakao and Viber Concepts: Tool Use Entities: OpenAI ChatGPTapp WhatsApp Kakao Source: twitter.com
• prio 6 Nobie launches an Excel-compatible runtime for agents and humans Concepts: Tool Use Agents Entities: Claude Gemini Source: nobie.com

💬 Opinions (13)

• prio 10 A Strix Halo home AI platform running 34 containers and local embedding/rerank workloads Concepts: RAG Embeddings Reranking Vector Database Entities: AMD Beelink Dify RAGFlow Source: habr.com
• prio 9 Shipping Mac and iOS apps without opening Xcode Concepts: Code Agents Tool Use Entities: Apple XcodeGen GitHub Homebrew Source: scottwillsey.com
• prio 9 Token pricing is not comparable without tokenizer counts Entities: Anthropic OpenAI Google xAI Source: playcode.io
• prio 8 Why better AI results start with the process, not the prompt Entities: Yandex ITMO Yandex Practicum Habr Source: habr.com
• prio 8 Go-Flavored Concurrency in C with pthreads Source: antonz.org
• prio 7 How a QA team automated Acceptance Criteria drafting with AI Concepts: MCP Tool Use Entities: Banki.ru Jira Confluence Figma Source: habr.com
• prio 7 Conceptual walkthrough of transformer tokenization and embeddings Concepts: Embeddings Source: habr.com
• prio 7 Control the ideas, not the code Entities: Redis DeepSeek X DeepSeek V4 Source: antirez.com
• prio 7 Using CPU affinity to keep local LLM workloads and VirtualBox on Intel P-cores in Windows 11 Entities: Intel Microsoft LM Studio VirtualBox Source: habr.com
• prio 7 Why tool mode does not guarantee protocol compliance Concepts: Tool Use Entities: Claude Gemini SLM Source: t.me
• prio 6 Why a local agent may become the main interface for software Concepts: Agents MCP Tool Use Entities: Yandex Source: habr.com
• prio 6 Vibe coding used to debug a Qlik Sense variable bug and build an audit script Concepts: Tool Use Entities: Qlik DAR KORUS Consulting QsAppMetadataConnector Source: habr.com
• prio 6 Why LLMs often ignore instructions Concepts: Agents Source: t.me

FAQ

What is in the 2026-07-13 AI brief?

The 2026-07-13 brief selected 104 signal items for AI builders and filtered 216 items as noise, using the radar’s community-relevance scoring.