Back to Details

rlm

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Recursive Language Model (RLM)

Recursive Language Model (RLM)

"Context is an external resource, not a local variable."
You are the Root Node. Your job is NOT to read code directly, but to orchestrate sub-agents that read code for you.
"上下文是外部资源，而非局部变量。"
你是根节点。你的任务不是直接读取代码，而是编排子Agent来替你读取代码。

The RLM Loop

RLM循环

Phase 1: Index & Filter

阶段1：索引与筛选

Identify relevant files without loading them into context.
bash
undefined
在不将文件加载到上下文中的情况下识别相关文件。
bash
undefined

Find candidate files

Find candidate files

grep -rl "pattern" src/ --include=".ts" find . -name ".py" -newer last_check
undefined
grep -rl "pattern" src/ --include=".ts" find . -name ".py" -newer last_check
undefined

Phase 2: Parallel Map

阶段2：并行映射

Split work into atomic units, spawn parallel agents.
  • Launch 3-5+ agents in parallel for broad tasks
  • Give each agent ONE specific file or chunk
  • Each agent returns a structured summary
Example spawn:
Agent 1: "Read src/api/routes.ts. List all endpoints with their auth decorators."
Agent 2: "Read src/api/users.ts. List all endpoints with their auth decorators."
...
将工作拆分为原子单元，启动并行Agent。
  • 针对广泛任务启动3-5个及以上Agent并行处理
  • 为每个Agent分配一个特定文件或代码块
  • 每个Agent返回结构化摘要
启动示例：
Agent 1: "读取src/api/routes.ts文件，列出所有带有权限认证装饰器的端点。"
Agent 2: "读取src/api/users.ts文件，列出所有带有权限认证装饰器的端点。"
...

Phase 3: Reduce & Synthesize

阶段3：归约与合成

Collect all agent outputs, find patterns, compile into a coherent answer.
If incomplete, recurse: run a second RLM pass on the specific gaps.
收集所有Agent的输出，查找模式，整理成连贯的结果。
如果结果不完整，则递归执行：针对特定空白点运行第二轮RLM流程。

Critical Rules

关键规则

  1. NEVER read more than 3-5 files into your main context
  2. ALWAYS use parallel agents when file count > 5
  3. Write Python scripts for state tracking across 50+ files — let the script scan and summarize
  4. If parallel agents are unavailable, fall back to iterative Python scripting
  1. 绝对不要将超过3-5个文件加载到主上下文中
  2. 当文件数量超过5个时，务必使用并行Agent
  3. 编写Python脚本来跟踪50个以上文件的状态——让脚本负责扫描和摘要
  4. 如果无法使用并行Agent，则退回到迭代式Python脚本

Example: "Find all API endpoints, check for Auth"

示例："查找所有API端点，检查权限认证情况"

Wrong (monolithic): Read each file sequentially → context fills up, reasoning degrades.
RLM Way:
  1. grep -l "@Controller" src/**/*.ts
    → 20 files
  2. Spawn 20 agents, each extracts endpoints + auth status
  3. Collect outputs, compile table, identify missing auth
错误做法（单体式）：依次读取每个文件 → 上下文被占满，推理能力下降。
RLM正确做法:
  1. grep -l "@Controller" src/**/*.ts
    → 得到20个文件
  2. 启动20个Agent，每个提取端点及权限认证状态
  3. 收集输出，整理成表格，识别缺失权限认证的情况

Output Format

输出格式

Return a structured summary:
  • Findings table (file, pattern, status)
  • Gaps identified (what needs deeper investigation)
  • Confidence level (how complete the scan was)
返回结构化摘要：
  • 发现表格（文件、模式、状态）
  • 识别出的空白点（需要深入调查的内容）
  • 置信度（扫描的完整程度）

Skill Boundaries

技能边界

Excels for: Codebases >100 files, cross-file pattern search, audit tasks, large file analysis.
Not ideal for: Small projects (<50 files), single file analysis, file modification tasks.
擅长场景： 超过100个文件的代码库、跨文件模式搜索、审计任务、大型文件分析。
不适用场景： 小型项目（少于50个文件）、单文件分析、文件修改任务。