A review of recent studies on chunking for Retrieval-Augmented Generation (RAG) finds no single best chunk size, with results varying by corpus and method. A chemistry study by Amiri and Bocklitz (202…
A June 2026 study by LightSpeed and Tencent found that four frontier models—Claude Sonnet 4.5, DeepSeek-V4-Pro, GLM-5, and Gemini-3-Flash—misjudged their total context size with median relative error …
A July 2026 paper (arXiv 2607.01935) identifies 'ghost memory' as a state coordination failure in agent memory systems, where old, current, and transition facts coexist and mislead the model, and repo…
A 2026 study by Daeyoung Roh and Donghee Han of 12,000 paired agent trajectories across HotpotQA, 2WikiMultiHopQA, and MuSiQue found that agentic RAG failures often stem from procedural errors: on MuS…
A July 2026 study of 44 language models found that asking for JSON output collapsed answer diversity: the modal answer rose from 41% to 64% of responses, distinct answers per prompt fell from 52 to 36…
Moonshot's Kimi K3, a 2.8-trillion-parameter mixture-of-experts model released July 16, achieves a 1M-token context window with decoding up to 6.3x faster than its predecessor by using Kimi Delta Atte…
OpenAI's GPT-5.6 Sol, generally available since July 9, 2026, introduces Ultra mode that decomposes tasks into parallel subagents with isolated context windows, lifting Terminal-Bench 2.1 scores from …
Wire, a provider of context containers for AI agents, is moving off Cloudflare Durable Objects due to four structural limits: the vector index lived outside the object, compute could not be placed nex…
AI agent reliability fails primarily because each run assembles context differently, with retrieval order, memory state, and tool outputs varying even at temperature zero, leading to accuracy variatio…
A 2026 study from UC San Diego, UC Berkeley, Texas A&M, and UIUC found that context pruning—masking stale observations from an AI agent's history—improves accuracy by up to 11.7 points in some setups …
Graph-based retrieval outperforms vector RAG on multi-hop reasoning and global summarization, with Microsoft's GraphRAG beating vector RAG on answer comprehensiveness 72 to 83 percent of the time and …
Context bloat, the accumulation of low-signal tool-call output in an agent's context window, degrades long-running agent performance. Anthropic's analysis found token usage explains 80% of performance…
Chunking strategies determine which slices of a document reach an AI model's context window, directly shaping retrieval accuracy. A 2026 benchmark found recursive 512-token chunking reached 69% accura…
A wave of research shows that AI agents which manage their own context window using learned policies outperform those relying on fixed truncation rules. Approaches like AgentSwing and budget-aware rei…
A 2026 University of British Columbia study found 21.8% structural waste across 4.45 billion tokens in LLM context windows, proposing demand paging as a solution. The researchers argue that treating t…
The Model Context Protocol's Enterprise-Managed Authorization extension, stable since June 18, 2026, redefines authorization as a context-scoping mechanism that determines which data sources an AI age…
The Model Context Protocol (MCP) will drop session state in its 2026 redesign, removing the initialize handshake and MCP-Session-Id header to enable stateless operation. The release candidate, publish…
MCP Tasks, introduced in the 2026-07-28 Model Context Protocol specification, allow servers to respond to tool calls with a durable task handle instead of a blocking result, enabling context offloadin…
A Peking University team presented Meta Context Engineering (MCE) at ICML 2026, achieving an 89.1% average improvement over a base agent across five domains, outperforming the previous best method (AC…
Anthropic's Claude Fable 5, released June 9, introduces safety classifiers that can decline requests mid-task and a fallback system that retries them on Claude Opus 4.8, making mid-conversation model …