什么是 Context Compressor？

Context Compressor 是一款专门设计的工具，用于处理因上下文过度堆积而导致性能下降的长周期会话。通过识别上下文何时接近极限，它能智能地总结历史交互，并将关键的项目状态转移到新会话中，确保您的 Openclaw Skills 始终保持敏锐且响应迅速。

该技能可以防止在复杂的代码迭代过程中出现常见的 AI 问题，如重复、细节丢失或“迷失在中间”现象。它在丢弃冗余聊天记录的同时，保留了关键决策、代码更改和待办任务，提供无缝衔接，在不增加臃肿上下文窗口负担的情况下保持连贯性。

下载入口:https://github.com/openclaw/skills/tree/main/skills/maddiedreese/context-compressor

安装与下载

1. ClawHub CLI

从源直接安装技能的最快方式。

2. 手动安装

将技能文件夹复制到以下位置之一

优先级：工作区 > 本地 > 内置

3. 提示词安装

将此提示词复制到 OpenClaw 即可自动安装。

Context Compressor 应用场景

• 管理具有大量迭代和改进的长周期编程会话。
• 防止在智能体开始出现重复或丢失项目细节追踪时性能下降。
• 在大型开发项目中，在达到硬性 Token 限制之前主动管理资源。
• 使用 Openclaw Skills 在多个会话交接中保持项目连贯性。

1. 该技能持续监控会话元数据，以跟踪实时上下文使用情况。
2. 当达到用户定义的阈值（默认为 80%）时，自动触发压缩工作流。
3. 提取高信号数据，包括关键架构决策、当前文件状态和待办事项。
4. 以压缩摘要为新基础，为智能体启动一个全新的会话。
5. 自动转移项目记忆，确保无缝的开发体验，且不丢失工作进度。

Context Compressor 配置指南

要在 Openclaw Skills 环境中配置 Context Compressor，请使用以下 CLI 命令：

# 设置压缩触发阈值（例如 75%）
context-compressor set-threshold 75

# 调整总结过程的深度
context-compressor set-depth detailed

# 检查当前的上下文状态和使用情况
context-compressor status

Context Compressor 数据架构与分类体系

该技能通过结构化的交接过程组织数据，生成特定的产出物以确保在 Openclaw Skills 会话之间的持久性：

Context Compressor Skill

Automated context management for long-running Clawdbot sessions. Detects when context limits approach, compresses old conversation history, and seamlessly transfers to a fresh session.

When to Use

• Long coding sessions with extensive context accumulation
• Projects with many iterations and refinements
• When noticing Claude repeating itself or losing track of details
• Proactively before hitting hard context limits
• During heartbeats or idle moments when user isn't actively waiting

How It Works

1. Monitoring: Continuously tracks context usage via session metadata
2. Compression: When threshold reached (configurable, default 80%), summarizes old messages
3. Preservation: Extracts key decisions, code changes, file states, and action items
4. Handoff: Initiates new session with compressed context as foundation
5. Continuity: User experiences seamless transition with all important context preserved

Key Features

• Smart Summarization: Preserves decisions, code, file states, not just raw text
• Configurable Thresholds: Set when compression triggers (70-90% of context limit)
• Background Operation: Works during heartbeats or low-activity periods
• Selective Retention: Keeps important files, decisions, TODOs; compresses chaff
• Session State Transfer: New session inherits all critical context automatically

Core Concepts

Context Degradation Patterns

As sessions grow long, watch for these signs:

• Repetitive responses ("As I mentioned earlier...")
• Missing references to earlier decisions
• Forgetting file modifications
• Asking to repeat information
• General coherence degradation

Compression Strategy

1. Extract Core Intelligence:

   • All decisions made and their rationale
   • File paths and their current state
   • Pending tasks and their status
   • Important code snippets or configurations
   • User preferences and patterns

2. Condense History:

   • Remove filler, backtracking, dead ends
   • Keep only high-signal turns
   • Merge related iterations into summaries
   • Preserve critical code snippets inline

3. Format for Efficiency:

   • Use compact representations
   • Reference files rather than dump contents
   • List decisions as bullet points
   • Include only relevant code context

Usage

Automatic Mode (Recommended)

The skill runs automatically during heartbeats and idle periods. Configure threshold:

# Set compression to trigger at 75% context usage
context-compressor set-threshold 75

# Check current status
context-compressor status

Manual Trigger

# Force compression and session reset
context-compressor compress --force

Configuration

# View all settings
context-compressor config

# Adjust summarization depth
context-compressor set-depth brief|detailed|comprehensive

# Set quiet hours (compression won't run)
context-compressor set-quiet-hours 23:00-07:00

Output

When compression occurs, the skill produces:

1. Summary File: memory/compressed-{session-id}.md

   • Executive summary of session
   • Key decisions made
   • Files modified and their states
   • Pending tasks
   • Code snippets worth preserving

2. Session Handoff: Automatic new session with:

   • User context (USER.md)
   • Project memory (MEMORY.md)
   • Compressed session summary
   • Current working state

Best Practices

1. Regular Compression: Don't wait for limits. Compress proactively every few hours
2. Preserve Code: Always keep actual code snippets, not just references
3. Track Decisions: Explicitly note WHY decisions were made, not just WHAT
4. Keep TODOs: Mark incomplete work clearly for continuity
5. Reference Files: Point to files rather than embedding full contents

Integration Points

• Heartbeats: Runs compression checks during heartbeat cycles
• Memory System: Writes compressed summaries to memory files
• Session Management: Coordinates with session spawn for handoff
• File Tracking: References current file states accurately

Technical Details

Compression Algorithm

1. Parse session transcript into atomic turns
2. Score each turn for importance (decisions = high, chat = low)
3. Keep top N% of turns by importance score
4. Summarize remaining turns into executive summary
5. Extract and preserve code blocks separately
6. Generate session transfer package

Thresholds

• Conservative (70%): Trigger early, preserve more context
• Balanced (80%): Default, good for most workflows
• Aggressive (90%): Push limits, maximum session length
• Manual Only: Disable auto-trigger, compress on demand

File References

The compressor tracks:

• Modified files and their paths
• Configuration changes
• New files created
• Deleted files
• Directory structure changes

Troubleshooting

Compression Too Frequent

# Increase threshold
context-compressor set-threshold 85

Context Lost After Handoff

Check that:

1. Compressed summary was generated (memory/compressed-*.md)
2. New session loaded memory files
3. Critical files weren't in the "chaff" that got dropped

Performance Impact

Compression runs in background and should complete in <30s for typical sessions. If slower:

• Reduce summarization depth
• Increase threshold to compress less frequently
• Exclude large files from compression scope

Examples

Typical Workflow

User: Working on notes app sidebar
[Session runs 2 hours, many iterations]

Heartbeat triggers → Context at 78% → Auto-compress → New session
User: (no interruption, seamless continuation)
New session has: executive summary, key decisions, file states, TODOs

Manual Recovery

User notices Claude repeating itself
User: "context-compressor compress --force"
Compressor summarizes → New session starts → User continues seamlessly

Related Skills

• memory-system: Underlying memory infrastructure
• self-improving-agent: Learns from session patterns
• sessions-spawn: Handles new session creation

See Also

• Context Engineering Skills Collection
• Research on "lost-in-the-middle" phenomenon in LLM context windows