Models and Algorithms
给 Agent 加了记忆，它反而变笨了
记忆几乎总是被当作纯增益功能加上去的，但 2026 年的多份公开测量显示：加了记忆的 Agent 在不少任务上比不加更差。本文拆解稀释、误差累积、陈旧状态三种退化机制，对比 compaction、结构化笔记、子 Agent 隔离三种手段的取舍，并给出写入侧的过滤与遗忘策略和三组可自测的对照实验。
Topic
Models and Algorithms
Model capabilities, assessment methods, and causal chains behind engineering choices.
Featured article
Latest articles
Models and Algorithms
Why the Same Question Produces Different Answers
Different answers to the same question are not always just randomness. This article separates generation, conversation context and version changes, and shows which differences need checking.
Models and Algorithms
Don’t Bet a Critical Workflow on One AI Tool
A practical framework for assessing dependence on one AI service through inputs, task rules, acceptance criteria, adapters, human takeover and tested degraded modes.
Models and Algorithms
Did Recommendations Actually Make Choosing Easier?
A close look at choice overload, ranking signals, exploration and agency to separate what recommendations save us from what they cannot decide for us.
Models and Algorithms
How AI Actually Improves Productivity at Work: By Cutting Waiting, Switching and Rework
Drawing on research in writing, customer support, workplace collaboration and software development, this article explains when AI speeds direct tasks, when it shifts time into verification and rework, and how teams should measure net efficiency.
Models and Algorithms
When AI Starts Working for You, the Real Change Is Not Efficiency
This article distinguishes task automation from workflow delegation, examines how continuous AI execution changes judgement, permissions, evidence, accountability and skill development, and proposes a practical five-step governance framework.
Models and Algorithms
The EU AI Act Enters Its Implementation Phase: An Enterprise Compliance Guide
As the EU AI Act enters its implementation phase, compliance is becoming a basic capability for entering the European market and operating AI there over time. This guide follows four arguments: responsibility boundaries, risk-based controls, evidence across the supply chain, and governance embedded in everyday operations.
Models and Algorithms
From 128K to 1M: What Long-Context Workflows Save and What They Add
Put long context back into concrete work such as contract review, cross-version comparison, and technical decisions. This article explains what 128K and 1M actually solve, and why retrieval, citations, and version control still matter.
Models and Algorithms
What Actually Makes AI Expensive?
From final training runs to HBM, advanced packaging, networking, depreciation, utilisation, inference scheduling and customer total cost of ownership, this article maps where AI spending goes and why a lower unit price does not guarantee lower total expenditure.