Kimi K3's 1M context runs mostly on linear attention
Kimi K3's 1M-token window runs on Kimi Delta Attention: hybrid linear attention with a fixed-size state. Why cheap long context still needs context curation.
Further reading
Articles about Context Window
28 articles from the Wire blog, sorted newest first. Return to the Context Window definition for context.
GPT-5.6 subagents: context isolation in the weights
GPT-5.6 Sol's Ultra mode gives each subagent its own context window. It solves context isolation but not handoff, and it multiplies your token bill fast.
Context pruning helps agents, until it doesn't
Context pruning helps AI agents in one regime and hurts in another. A 2026 study of models from 4B to 284B maps when to prune stale context and when not to.
Context bloat: why long-running agents break
Context bloat is when accumulated tool-call output crowds out an agent's task. Tool calls, not window size, break long-running agents. Here is the fix.
How agents manage their own context window
Agent context management is shifting from fixed harness rules to learned, runtime decisions an agent makes about its own window. What the 2026 research shows.
Demand paging for the AI context window
A 2026 systems paper found 21.8% of tokens in agent context windows are wasted. Demand paging treats the AI context window as L1 cache, not full memory.
MCP authorization decides what context agents see
MCP authorization became a context control plane in 2026. RFC 8707 token scoping decides which sources an agent can ever pull into its own context window.
MCP Tasks: long-running work as context offloading
MCP Tasks let a server return a durable handle instead of a blocking result, keeping a long-running tool call's interim state off the agent's context window.
Sub-agent context isolation: the fix for context rot
Sub-agent context isolation gives each agent its own scoped window, stopping the context rot that kills multi-agent runs. Here's the pattern and its limits.
Claude Opus 4.8 hallucinates less by answering less
Claude Opus 4.8 tops a hallucination benchmark without getting more accurate. It learned to abstain. Why retrieval honesty is a context engineering win.
RAG, long context, or fine-tuning: how to choose
RAG, long context, or fine-tuning? A 2026 decision guide on cost, accuracy, and freshness, with a use-case table for choosing the right one in production.
File-native agents: when reading files backfires
A 9,649-experiment study found file-native retrieval lifts frontier-model accuracy 2.7% and drops open-source accuracy 7.7%. Match architecture to the model.
Context offloading: 3 patterns for AI agents
Context offloading keeps an AI agent's working context window small by moving state to a destination outside it. Three patterns, and what each one costs.
MCP vs Skills vs CLI: which one wastes the least context?
GitHub's MCP costs tens of thousands of tokens before any work begins. We compare MCP, Claude Skills, and CLI by context cost, not by user preference.
Tool calling is retrieval: what Needle proves at 26M
A 26M-parameter model just matched Gemini at function calling. Here is what Needle's distillation result means for MCP and agent context engineering.
Long context tripled hallucinations in 35 open models
A 172-billion-token study across 35 open models found hallucination rates triple from 32K to 128K context, and exceed 10% at 200K for every model tested.
Progressive tool loading is the new MCP context pattern
Preloading every MCP tool into an agent's context is the bottleneck of 2026. Progressive tool loading defers definitions until needed and saves tokens.
Why token cost doesn't scale with knowledge base size
AI token usage scales with knowledge base size only when the full corpus loads per query. The real variable is selective context delivery, not KB size.
Context budgets: how to allocate tokens for AI agents
A practical guide to context budgets for AI agents. How to allocate tokens across system prompts, tools, retrieval, history, and a buffer in production.
RAG vs long context: what the 2026 data shows
RAG vs long context in 2026: which wins on cost, speed, and accuracy, and when each one beats the other in production. What the benchmarks actually show.
Context compression: why less context means better AI
Context compression reduces AI agent memory usage by 26-54% while preserving task performance. Here's how it works and why bigger context windows aren't the answer.
Why AI agents forget mid-task (and how to fix it)
65% of agent failures come from context drift, not token limits. Here's how context compression keeps long-running AI agents on track.
Why AI Agent Memory Keeps Failing
AI agent memory fails because it's a context engineering problem, not a storage problem. Research reveals three failure modes and what actually works.
Why AI Coding Assistants Can't See Your Codebase
84% of developers use AI coding tools, but only 29% trust the output. The problem has less to do with models and more to do with codebase context.
What Is a Context Window?
A context window is the total text an AI model can process at once. Learn how they work, why size isn't everything, and what actually affects performance.
Context engineering: what replaces prompt engineering
Prompt engineering has a new successor: context engineering. Learn why Karpathy and Tobi Lütke made the switch, and what it means for production AI systems.
Why Does ChatGPT Forget Everything?
AI doesn't forget because it's broken — it forgets because everything gets crammed into one place. Here's the technical explanation and how to fix it.
Context Rot: Why AI Performance Degrades With More Information
Research shows LLMs drop from 95% to 60% accuracy as context grows stale. Here's how context rot degrades AI performance and why bigger windows won't help.
Put context into practice
Create your first context container and connect it to your AI tools in minutes.
Create Your First Container