Skip to main content

Context

注意力计算复杂度是 O(n2)O(n^2), 且是稀疏的 (模型不会均匀地关注所有输入):

  • 上下文窗口受硬件边界限制.
  • 有效上下文小于标称上下文: coding agent 只能有效利用其中的 10-15 %\%.
  • Lost in the middle: 中间内容容易被忽略, 更关注开头和结尾.

Engineering​

LLM 并未统一利用其上下文, 它们的准确性和可靠性会随着输入令牌数量的增加而下降, 称之为上下文腐烂 (Context Rot).

因此, 仅仅在模型的上下文中拥有相关信息是不够的: 信息的呈现方式对性能有显著影响. 这凸显了 上下文工程 的必要性, 优化相关信息的数量并最小化不相关上下文以实现可靠的性能. e.g. custom gemini CLI command.

Context Engineering

Planning with Files

Manus:

  1. Design around KV-cache:
    • 稳定内容放前面: system prompt, tool definitions.
    • 动态内容放后面: chat history, user input.
    • 避免在稳定前缀中插入可变内容: e.g. 禁止在 system prompt 中插入时间戳.
  2. Plan is required
  3. Files are memory
  4. Don't get few-shotted: get rid of repetitive actions
  5. Manipulate attention through recitation
Start of context: [Original goal - far away, forgotten]
...many tool calls...
End of context: [Recently read task_plan.md - gets ATTENTION!]

Dynamic Discovery​

Dynamic context discovery:

  • 工具响应 -> 文件.
  • 终端会话 -> 文件.
  • 上下文压缩时引用对话历史.
  • 按需加载.
  • 渐进式披露.

Personalization​

Meta-prompting for memory extraction:

You are a [USE CASE] agent whose goal is [GOAL].
What information would be important to keep in working memory during a single session?
List both fixed attributes (always needed) and inferred attributes (derived from user behavior or context).
Memory System

Memory system:

  • Repeatable memory loop: inject → reason → distill → consolidate.
  • Enforce precedence: current user message > session context > memory.