Skip to content

🛰 AI Brief — Nov 20, 2026

🥇 GrepMem: Lossless, Index-Free Memory for Long-Horizon LLM Agents · prio 13

For builders designing long-horizon AI agents, this research challenges complex embedding-based vector memory architectures by showing that a simple append-only text log paired with regex search can achieve higher recall accuracy with lower infrastructure overhead. It directly addresses community knowledge gaps around agent memory and context engineering by offering a practical, lossless alternative to lossy summarization. Concepts: Agent Memory Agents Context Engineering Tool Use Entities: Hugging Face CORE Connected Papers Litmaps DagsHub Qwen3.5-9B 2 sources: arxiv.org, arxiv.org

🥈 CompAct-Bench: Evaluating Working-Context Compaction in Long-Horizon Agent Tasks · prio 11

For builders running long-horizon agents or coding assistants that compress conversation history to manage token limits, this paper highlights that context compaction frequently degrades downstream task accuracy and distorts task state even in frontier models. Understanding these failure modes helps teams design more reliable context-management and verification mechanisms instead of relying naively on LLM summarization. Concepts: Context Engineering Agents LLM Evals Entities: Claude GPT Source: arxiv.org

🥉 alphaXiv Releases OpenResearch, an Open-Source Local Harness for Research Agents · prio 11

OpenResearch provides a local workspace and execution harness directly compatible with the community’s primary coding agents (Claude Code, Cursor, Codex, and Google Antigravity). It demonstrates a practical pattern for agent memory by persisting experiment history, logs, and artifacts in a local SQL database so agents can build on previous trials. Concepts: Agents Agent Memory Code Agents Entities: alphaXiv Google Hugging Face Modal Source: github.com

4️⃣ DriftGuard: Online Memory-Drift Detection via Causal Attribution for LLM Agents · prio 10

For builders creating multi-step agents that rely on persistent memory and tool calling, memory poisoning or drift poses a critical reliability challenge. DriftGuard provides a concrete causal method to detect and filter harmful retrieved memories at serving time without incurring prohibitive rollout latency. Concepts: Agent Memory Agents Tool Use Source: anonymous.4open.science

5️⃣ AutoRAGBench: A Benchmark for LLM Agents Automating RAG System Development · prio 9

AutoRAGBench evaluates whether coding agents can automate end-to-end RAG architecture design and iterative optimization, directly connecting to the community’s interest in autonomous coding agents and weak area in RAG development. The findings demonstrate that agents like GPT-6.1 Sol paired with Codex CLI can iterate on and surpass human-designed baseline RAG pipelines. Concepts: RAG Agents Code Agents LLM Evals Entities: arXiv GPT-6.1 Sol 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 · Context Engineering

FAQ

What is in the 2026-11-20 AI brief?

The 2026-11-20 brief selected 50 signal items for AI builders and filtered 199 items as noise, using the radar’s community-relevance scoring.