rlm-context-manager

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

RLM Context Manager

RLM 上下文管理器

Purpose: Enable Claude to process documents and contexts that exceed typical context window limits, matching Gemini's large context capabilities through intelligent chunking and recursive processing.

用途：让Claude能够处理超出常规上下文窗口限制的文档和上下文，通过智能分块和递归处理，达到与Gemini相当的大上下文处理能力。

Architecture (RLM Paper Implementation)

架构（RLM论文实现）

Based on arXiv:2512.24601 - Recursive Language Models by Zhang, Kraska, Khattab (MIT CSAIL).

Component	IRP Implementation	Model
Root LLM	Main Claude Code conversation	Claude Opus 4.5
Sub-LLM ( `llm_query` )	`rlm-subcall` Task agent	Claude Haiku
External Environment	Persistent Python REPL	Python 3
State Persistence	`${SKILLS_ROOT}/rlm-context-manager/state/`	Pickle

基于论文arXiv:2512.24601 - 由MIT CSAIL的Zhang、Kraska、Khattab提出的递归语言模型。

组件	IRP实现方式	模型
根LLM	主Claude对话	Claude Opus 4.5
子LLM（ `llm_query` ）	`rlm-subcall` 任务Agent	Claude Haiku
外部环境	持久化Python REPL	Python 3
状态持久化	`${SKILLS_ROOT}/rlm-context-manager/state/`	Pickle

Use Cases

适用场景

Processing large codebases for analysis
Analyzing lengthy documents (research papers, legal docs, logs)
Working with Mnemosyne ledger archives
Cross-session context restoration
Bridging context between Claude and Gemini

处理大型代码库以进行分析
分析长文档（研究论文、法律文档、日志）
处理Mnemosyne账本归档文件
跨会话上下文恢复
实现Claude与Gemini之间的上下文互通

Commands

命令

/rlm init <context_path>     - Initialize REPL with large context file
/rlm status                  - Show current RLM state (chars loaded, chunks, buffers)
/rlm query <question>        - Query the loaded context
/rlm chunk                   - Materialize context into chunk files
/rlm synthesize              - Merge collected evidence into final answer
/rlm reset                   - Clear RLM state
/rlm export                  - Export buffers to file

/rlm init <context_path>     - 初始化REPL并加载大上下文文件
/rlm status                  - 显示当前RLM状态（已加载字符数、分块数、缓冲区）
/rlm query <question>        - 对已加载的上下文进行查询
/rlm chunk                   - 将上下文生成分块文件
/rlm synthesize              - 将收集到的信息合并为最终答案
/rlm reset                   - 清除RLM状态
/rlm export                  - 将缓冲区内容导出到文件

Quick Start

快速开始

bash

undefined

bash

undefined

1. Initialize with a large context file

1. 使用大上下文文件初始化

python3 ${SKILLS_ROOT}/rlm-context-manager/scripts/rlm_repl.py init /path/to/large_document.txt

2. Check status

2. 查看状态

python3 ${SKILLS_ROOT}/rlm-context-manager/scripts/rlm_repl.py status

3. Scout the context (peek at beginning and end)

3. 预览上下文（查看开头和结尾部分）

python3 ${SKILLS_ROOT}/rlm-context-manager/scripts/rlm_repl.py exec -c "print(peek(0, 3000))"

4. Create chunks for sub-LLM processing

4. 生成分块以供子LLM处理

python3 ${SKILLS_ROOT}/rlm-context-manager/scripts/rlm_repl.py exec <<'PY' paths = write_chunks('${SKILLS_ROOT}/rlm-context-manager/state/chunks', size=200000, overlap=0) print(f"Created {len(paths)} chunks") PY

undefined

undefined

Step-by-Step Procedure

分步操作流程

1. Initialize the REPL State

1. 初始化REPL状态

bash

python3 ${SKILLS_ROOT}/rlm-context-manager/scripts/rlm_repl.py init <context_path>
python3 ${SKILLS_ROOT}/rlm-context-manager/scripts/rlm_repl.py status

bash

python3 ${SKILLS_ROOT}/rlm-context-manager/scripts/rlm_repl.py init <context_path>
python3 ${SKILLS_ROOT}/rlm-context-manager/scripts/rlm_repl.py status

2. Scout the Context

2. 预览上下文

bash

undefined

bash

undefined

Peek at beginning

查看开头部分

python3 ${SKILLS_ROOT}/rlm-context-manager/scripts/rlm_repl.py exec -c "print(peek(0, 3000))"

Peek at end

查看结尾部分

python3 ${SKILLS_ROOT}/rlm-context-manager/scripts/rlm_repl.py exec -c "print(peek(len(content)-3000, len(content)))"

Search for patterns

搜索指定模式

python3 ${SKILLS_ROOT}/rlm-context-manager/scripts/rlm_repl.py exec -c "print(grep('pattern', max_matches=10))"

undefined

python3 ${SKILLS_ROOT}/rlm-context-manager/scripts/rlm_repl.py exec -c "print(grep('pattern', max_matches=10))"

undefined

3. Choose Chunking Strategy

3. 选择分块策略

Semantic chunking: For structured formats (markdown headings, JSON, log timestamps)
Character chunking: Default ~200k chars with optional overlap

语义分块：适用于结构化格式（Markdown标题、JSON、日志时间戳）
字符分块：默认约20万字符，可设置重叠部分

4. Materialize Chunks

4. 生成分块文件

bash

python3 ${SKILLS_ROOT}/rlm-context-manager/scripts/rlm_repl.py exec <<'PY'
paths = write_chunks('${SKILLS_ROOT}/rlm-context-manager/state/chunks', size=200000, overlap=0)
print(len(paths))
print(paths[:5])
PY

bash

python3 ${SKILLS_ROOT}/rlm-context-manager/scripts/rlm_repl.py exec <<'PY'
paths = write_chunks('${SKILLS_ROOT}/rlm-context-manager/state/chunks', size=200000, overlap=0)
print(len(paths))
print(paths[:5])
PY

5. Sub-LLM Processing Loop

5. 子LLM处理循环

For each chunk, invoke the rlm-subcall subagent:

Task: rlm-subcall
Prompt: "Query: <user_query>. Chunk file: <chunk_path>. Extract relevant information."
Model: haiku

The sub-LLM returns structured JSON:

json

{
  "chunk_id": "...",
  "relevant": [{"point": "...", "evidence": "...", "confidence": "high|medium|low"}],
  "missing": ["..."],
  "suggested_next_queries": ["..."],
  "answer_if_complete": "..."
}

对每个分块，调用rlm-subcall子Agent：

Task: rlm-subcall
Prompt: "Query: <user_query>. Chunk file: <chunk_path>. Extract relevant information."
Model: haiku

子LLM会返回结构化JSON：

json

{
  "chunk_id": "...",
  "relevant": [{"point": "...", "evidence": "...", "confidence": "high|medium|low"}],
  "missing": ["..."],
  "suggested_next_queries": ["..."],
  "answer_if_complete": "..."
}

6. Synthesis

6. 结果合成

Collect sub-LLM outputs and synthesize final answer in main conversation.

收集子LLM的输出结果，在主对话中合成最终答案。

REPL Helper Functions

REPL辅助函数

Function	Description
`peek(start, end)`	Return substring of context
`grep(pattern, max_matches, window)`	Regex search with context window
`chunk_indices(size, overlap)`	Calculate chunk boundaries
`write_chunks(out_dir, size, overlap)`	Materialize chunks to files
`add_buffer(text)`	Store intermediate results

函数	描述
`peek(start, end)`	返回上下文的子字符串
`grep(pattern, max_matches, window)`	带上下文窗口的正则搜索
`chunk_indices(size, overlap)`	计算分块边界
`write_chunks(out_dir, size, overlap)`	将分块生成为文件
`add_buffer(text)`	存储中间结果

Guardrails

防护规则

Do NOT paste large raw chunks into main chat context
Use REPL to locate exact excerpts; quote only what you need
Subagents cannot spawn other subagents (orchestration stays in main conversation)

Keep scratch/state files under

${SKILLS_ROOT}/rlm-context-manager/state/

Prefer chunk sizes ~100k-300k chars per subagent call

请勿将大型原始分块粘贴到主聊天上下文中
使用REPL定位精确片段，仅引用所需内容
子Agent无法生成其他子Agent（编排逻辑保留在主对话中）
临时文件/状态文件请存放在
```
${SKILLS_ROOT}/rlm-context-manager/state/
```
目录下
每次调用子Agent时，分块大小建议为10万-30万字符

Integration with IRP Protocols

与IRP协议的集成

Mnemosyne Ledger

Mnemosyne账本

Use RLM to process large Mnemosyne archives
Export buffers as Mnemosyne packets for cross-session persistence

使用RLM处理大型Mnemosyne归档文件
将缓冲区内容导出为Mnemosyne数据包，实现跨会话持久化

Cross-Model (Gemini Bridging)

跨模型（Gemini桥接）

Load Gemini's large context exports
Chunk and analyze for Claude consumption
Synthesize into Mnemosyne-compatible format

加载Gemini的大上下文导出文件
分块后供Claude分析处理
合成为Mnemosyne兼容格式

CRTP Packets

CRTP数据包

RLM outputs can be formatted as CRTP packets for multi-model relay

RLM输出可格式化为CRTP数据包，用于多模型中继

File Structure

文件结构

rlm-context-manager/
├── SKILL.md              # This file
├── scripts/
│   └── rlm_repl.py       # Persistent Python REPL
├── agents/
│   └── rlm-subcall.md    # Sub-LLM agent definition
└── state/                # Runtime state (gitignored)
    ├── state.pkl         # Persisted REPL state
    └── chunks/           # Materialized chunk files

rlm-context-manager/
├── SKILL.md              # 本文档
├── scripts/
│   └── rlm_repl.py       # 持久化Python REPL
├── agents/
│   └── rlm-subcall.md    # 子LLM Agent定义
└── state/                # 运行时状态（已加入git忽略）
    ├── state.pkl         # 持久化REPL状态
    └── chunks/           # 生成分块文件

Credits

致谢

Based on the RLM paper by Zhang, Kraska, Khattab (MIT CSAIL) and the Claude Code RLM implementation by Brainqub3.

基于MIT CSAIL的Zhang、Kraska、Khattab所著的RLM论文，以及Brainqub3实现的Claude Code RLM版本。