Context Engineering 2.0: The Context of Context Engineering
A 30-year evolution from context-awareness to context-cooperation. Context Engineering is an entropy-reduction process aimed at bridging the cognitive gap between humans and machines.
30-Year Timeline
- 1994: Bill Schilit first proposed the concept of "context-aware computing"
- 2000: Anind Dey's team developed the Context Toolkit framework
- 2001: Anind Dey gave the still widely cited definition of context
- Era 1.0 — Context-Aware: Machines are like infants, only able to "eat baby food" (structured data)
- Era 2.0 — Context-Cooperative: Machines are like adults, able to "eat steak directly" (raw information)
- Perception upgrade: from single sensors to multimodal fusion
- Environmental sensing: location, identity, activity, etc.
- Layered memory architecture: short-term and long-term memory
- Sub-agent context isolation: e.g., Claude Code spawns sub-agents for independent tasks
- Lightweight references: store large files externally and keep only "pointers" in context
- From "passive response" to "active collaboration"
- Improved ability to consume high-entropy context: from "eating only processed food" to "digesting raw ingredients"
- Without Self-Baking: AI only recalls ("What did you say last time?")
- With Self-Baking: AI accumulates knowledge ("I know what you like")
- GitHub: https://github.com/GAIR-NLP/Context-Engineering-2.0
- Paper (PDF): https://arxiv.org/pdf/2510.26493
- Slides: https://pan.quark.cn/s/52ad1ae833a7
The Cognitive Gap
> Cognitive Gap = Human context-processing capability − Machine context-processing capability
Context Engineering collects, manages, and uses context information to pre-process high-entropy human intent and environmental states into low-entropy representations that machines can understand.
Context Engineering = an entropy-reduction process
Systematic Framework
Context Engineering = Collection × Management × Usage
Collection (how to gather context)
Management (how to manage context)
Usage (how to use context)
From Perception to Cooperation
| Era 1.0 | Era 2.0 | | --- | --- | | Passive response | Active collaboration | | Structured data only | Handles raw information | | Single sensor | Multimodal fusion |
Example of an Era 2.0 cooperative system: while you write a paper, the system analyzes your progress, notices you are stuck on chapter 3, and proactively asks "Want me to help organize your logic?" — after you agree, it generates an outline, you edit it, and it adapts based on feedback.
Self-Baking: From Quantitative to Qualitative Change
The essence of Self-Baking is separating "storage" from "learning":
This is the watershed from "tool" to "partner" — context engineering must undergo a qualitative shift, with core summaries fused into model parameters.