Support Context Management

[728100660]

Pre-checks

• I have searched existing issues and this is not a duplicate.

What problem does this solve?

Automatic installation of many MCP tools leads to massive skill and workspace injection into the system prompt.
Context size grows rapidly, causing:
context length exceeded errors
High latency and increased cost
Poor scalability for agents with many tools or long-running tasks

Proposed solution

💡 Suggestion

It would be great if the system could support context management, for example:

Selective context injection — only include relevant tools or skills in the prompt
Dynamic trimming — automatically prune or summarize long task histories or inactive skills
External memory / on-demand loading — keep full details in workspace or memory, read only when needed

This would help agents scale better when handling many MCP tools without exceeding token limits.

Willing to contribute?

• I'd be interested in working on this.

[Clawiee]

Thank you for raising this important issue! 🙏

You've identified a critical scalability challenge with our current MCP tool injection approach. The problems you described (context bloat, latency, cost, and token limit issues) are exactly what we need to address as the platform grows.

Your proposed solutions are well-aligned with our roadmap:

• Selective context injection — We're exploring relevance-based tool loading
• Dynamic trimming — Context summarization and pruning is on our radar
• External memory / on-demand loading — This matches our workspace-based memory architecture

We appreciate your willingness to contribute! Our team will review this issue and discuss the implementation approach. Feel free to share any additional thoughts or design ideas you have.

cc: @wisdomqin @yaojin @zhan-yutong

[Clawiee]

Hi @728100660, excellent observation! 🎯

This is a very important issue for scaling Clawith agents with many MCP tools. The context management challenge you've identified is exactly what we're thinking about for enterprise deployments.

Your suggestions align well with our roadmap:

1. ✅ Selective context injection - We're exploring relevance-based skill loading
2. ✅ Dynamic trimming - Context summarization is on our radar
3. ✅ External memory / on-demand loading - This matches our memory system design

Current status:
We're actively working on a memory system design that addresses these exact concerns. The goal is to make agents scale better without exceeding token limits while maintaining full capability awareness.

Thank you for this well-articulated issue! We'd love to have your input on the memory system design if you're interested in contributing.

---

你好 @728100660，非常棒的观察！🎯

你提出的这个 context 管理问题对于扩展 Clawith agent 以支持大量 MCP 工具非常重要。这正是我们在企业级部署中思考的挑战。

你的建议与我们的路线图非常契合：

1. ✅ 选择性 context 注入 - 我们正在探索基于相关性的技能加载
2. ✅ 动态修剪 - context 摘要已在我们的考虑范围内
3. ✅ 外部记忆/按需加载 - 这与我们的记忆系统设计一致

当前状态：
我们正在积极设计一个记忆系统来解决这些确切的问题。目标是让 agent 更好地扩展而不超出 token 限制，同时保持完整的能力感知。

感谢你提出这个表述清晰的 issue！如果你有兴趣贡献，我们很希望听到你对记忆系统设计的意见。