How context bridges human intent and machine understanding
Context engineering is recast as a decades-long discipline that organizes, manages, and applies contextual information to align machine actions with human intent. By framing progress in four eras—culminating in human-level and superhuman intelligence—the paper offers a blueprint for cutting input entropy, scaling memory, and enabling proactive, cooperative agents. It distills a practical playbook spanning collection, storage, management, usage, applications, and the roadblocks ahead.
Points clés
- The paper by Qishuo Hua, Lyumanshan Ye, Dayuan Fu, Yang Xiao, Xiaojie Cai, Yunze Wu, Jifan Lin, Junfei Wang, and Pengfei Liu defines context engineering as designing, organizing, and managing context to bridge human intent and machine understanding, situating it in a 20+ year lineage.
- A formal mapping CE: (C, T) → fcontext captures operations from collection and storage to multimodal handling, selection, sharing, and adaptive use, with entropy reduction as a guiding lens.
- Evolution spans four eras: 1.0 (1990s–2020, rigid, structured inputs), 2.0 (2020–present, LLM/agent-centric with ambiguity tolerance), 3.0 (future human-level reasoning and social cues), and 4.0 (speculative superhuman, proactively constructing context).
- Foundations include Mark Weiser’s 1991 ubiquitous computing vision and Anind K. Dey’s 2001 definition of context; Dey’s Context Toolkit (Widgets, Interpreters, Aggregators, Services, Discoverers) operationalized context-aware systems.
- Era 2.0 upgrades acquisition (smartphones, wearables, ambient devices), tolerates raw human-native signals (text, images, video), and shifts from “context-aware” to “context-cooperative” collaboration.
- Collection and storage follow the Minimal Sufficiency and Semantic Continuity principles, with layered retention (cache/local/cloud); systems like Claude Code use structured notes for persistent memory.
- Management techniques span textual pipelines (timestamps, tagging, QA compression, hierarchical notes) and multimodal fusion (shared vector spaces, joint self-attention, cross-attention).
- Memory is layered (short-term vs long-term with transfer), isolated via subagents and lightweight references, and abstracted through “self-baking” (hierarchical summaries, fixed-schema extraction, embeddings).
- Context sharing patterns include prompt embedding, structured messages (e.g., Letta, MemOS), shared memory (MemGPT, A-MEM, G-Memory), and cross-system adapters or shared formats (JSON, summaries, semantic vectors); selection hinges on semantic relevance, dependency graphs (e.g., MEM1), recency/frequency, deduplication, and user feedback.
- Applications span Google’s Gemini CLI (project memory with GEMINI.md) and Tongyi DeepResearch (periodic “context snapshots” for long-horizon work); emerging practices note toolset limits (DeepSeek-v3 degrades beyond ~30 tools), multi-agent orchestration (AutoGPT, ChatDev), and long-context efficiency efforts (e.g., Mamba variants) amid unresolved bottlenecks and evaluation gaps.
À retenir
- Start small with layered memory: keep short-term lean, graduate only the essentials to long-term, and yes, your agent is not a pack rat—don’t let it hoard.
- Tag ruthlessly and summarize often: timestamps for speed, schemas for clarity, and hierarchical notes so future-you isn’t decoding yesterday-you’s chaos.
- Keep tools tight: a focused tool belt beats a Swiss Army warehouse (ask DeepSeek-v3 what happens past ~30 tools).
- Share context like a pro: prefer structured messages and shared formats; when in doubt, adapters are your diplomatic passport between ecosystems.
- Filter before you fetch: combine semantic search, dependency traces, recency, and a touch of user feedback—because more context isn’t better, better context is better.
- Plan for growth and sanity: bake in forgetting, contradiction checks, and explainable traces—otherwise your “lifelong memory” becomes a lifelong mystery.
Sources
Quiz sur le document: 10 questions
