2026-10-24

Raw-RAG: dense retrieval and reranking over raw conversation history lead the compared memory systems in 6 of 7 benchmark-model settings, by 0.13-5.53 points over the strongest agentic-memory baseline. Read it for evidence preservation before adding memory complexity; the abstract leaves costs and token budgets unspecified. → Raw-RAG — raw-history retrieval as a memory baseline (source)

Agent Mosaic: on benign queries, GPT-5.6-Sol discloses prohibited cross-document conclusions in 7.08% of simulated answers (95% CI: 3.75-10.67%). Read it for the distinction between document access and permitted inference; these synthetic results do not estimate production leak rates. → Agent Mosaic — confidentiality leaks through combined evidence (source)

• Moonshot's internal cyber evaluation drops from 90.5% to 81.0% pass@3 after filtering cyber-related training data and RL environments. No hard timeout or output-token limit is applied; relative reduction does not establish low absolute capability. → Open-Weights Policy Fight (source)

• Kimi K3.1 Pro and two research notes: Raw-RAG — raw-history retrieval as a memory baseline, Agent Mosaic — confidentiality leaks through combined evidence. (model) (memory) (confidentiality)

• Agent Runtime Containment now separates refusal frequency from output confidentiality: the study finds refusals can still reveal the forbidden conclusion. (source)