Tag
讨论指出 MCP 工具定义会占用大量常驻上下文——有用户卸载一个大型 MCP server 后上下文从约 200K 降至约 10K tokens。文章提供了如何在首次对话前检查工具开销的实用建议，包括检查请求日志、按需禁用 server 及使用延迟加载。
Qwen4 models are rumored to feature a 1 million token context window, with Qwen4-max reportedly launching around October 1st and a possible Qwen4-27B announcement at Apsara Netherlands.
Baoyu recommends that Claude Code users run /autocompact 300k to lower the automatic compaction threshold from the default of approximately 967K to 300K, keeping the conversation history concise and significantly reducing token consumption in long conversations. Lydia Hallie adds that auto-compact replaces the entire conversation with a brief summary.
一名用户详细拆解了自己一周内使用 Grok Bot 的费用：升级到每月 $200 套餐，其中 52% 的成本来自 16 步以上的长任务，90% 的 token 是对已有上下文的重复读取。文章给出了测量用量、脚本化重复任务、将长任务交给独立代理等省钱建议。
文章讨论了在每个回合中把所有 MCP 工具 schema 塞进上下文是一种 harness 缺陷，会导致选错工具和 token 浪费；业界正在收敛到渐进式发现（progressive discovery）方案，如 Claude Code 的 MCP Tool Search 和 OpenAI 的延迟加载工具搜索，可减少 85% 以上的工具定义开销。
一条 Thread Reader 聚合帖宣称 GPT 6.1 Sol 默认上下文窗口达到 500k，但未提供官方来源佐证，可能属于传闻或推测性消息。
OpenAI 宣布 GPT-6.1 Sol 的默认上下文窗口从 272k 提升至 500k（约 80% 增幅），在 Codex 和 ChatGPT Work 中免费提供，可减少长任务中的上下文压缩次数并加快端到端完成速度。
作者分享用四个长期运行的 Claude Code 智能体运营自己的 AI 通讯的经验，指出最大瓶颈不是记忆或工具，而是上下文窗口，并介绍了智能体自写检查点、自清空、自我恢复的机制，以及相关的失败案例和开源工具。
OpenAI's Tibo (@thsottiaux) announced that the default context length for GPT-6.1 Sol has been increased to 500k tokens when used via Codex and Work, up from 272k (~80% increase), with plans to go higher to reduce compactions on context-heavy tasks.
A practitioner describes how their AI voice agent hallucinated a delivery date and then doubled down on it across multiple turns as its own output accumulated in context, and how separating tool-returned facts from the agent's own utterances fixed the issue more reliably than prompting the model to treat its answers as provisional.
一位推主根据 Anthropic 工程师 Lydia Hallie 的官方解释，澄清了 Claude Code 的 auto-compact 机制：默认在接近 967K 上下文时才触发压缩，导致长任务持续携带巨量上下文而快速消耗 5 小时配额。文中建议通过 /autocompact 400k、autoCompactWindow、promptCacheTtl 等配置提前压缩并优化缓存。
LangChain Academy is promoting its free "Foundation: Introduction to Deep Agents" course, a 46-lesson curriculum covering agent skills, context window protection, MCP, checkpointers, HITL, and subagent delegation for building long-running AI agents.
A self-run experiment compares a six-model 1-2B local swarm (modeled on a 1930s Soviet design institute's discipline-based org chart) against a single model on a document extraction task, finding the real win came from splitting inputs into blocks rather than agent collaboration (+0.05 from the swarm, with strict emergence at zero).
Periscope introduces a training-free inference method that arranges text chunks on a K×K grid and queries a frozen language model over local and strided spans to recover answers beyond the context window, matching or beating full-window reads on LongBench v2 and InfiniteBench at a fraction of the cost. The approach lets a 27B model read 4.5M-token contexts on a single 80GB GPU by replacing long reads with an evidence map.
A same-harness benchmark on a Mac Studio M5 Ultra 256GB compares Qwen3.8-Flash-Next (182GB oQ8e) vs Laguna-S-2.1 GGUF at up to 262K context. Qwen sustains ~4,200 tok/s linear prefill and stable decode, while Laguna degrades superlinearly; quality is a 4/4 draw with very different answer styles.
Rohan Paul explains why AI-generated fiction reads like slop — single-prompt chatbot design, limited context windows, and no memory of long-running story promises — and highlights Sherpa, which uses specialized agents (ideas, structure, research, prose, review) plus a Narrative World Model to track story history and continuity.
Memorizon decouples long-span supervision from attention in streaming world models by having each scored chunk retrieve a shared bounded top-K latent bank via camera co-visibility, enabling consistent revisits beyond the context window at minimal extra cost. Retrieval improves revisit consistency over a sliding-window baseline by up to 30%, and results show the model genuinely uses the retrieved latents.
Running a Discord agent powered by OpenClaw and GLM 5.3 Flash for 23 hours led to failures at around 65% of the 1M context window, exposing issues with caching and model performance under heavy use.
A technical report for a 10M context window model has been published by Voltropy, potentially revolutionizing AI's capacity to handle extremely long texts.
该论文研究语言模型自注意力机制的实际检索能力，通过仅保留每层每头注意力权重最高的令牌，估计在给定损失容忍度下保持性能所需的有效注意力集大小，发现较小集即可接近全注意力基线，且注意力集大小随上下文增长而增加，即使没有更多信息需要检索。