Context Management Strategies

Module 4, Lecture 4.2 | Section 2: Working with LLMs in Practice

When a context budget threshold is crossed, you need a strategy. This lecture covers three — sliding window, selective preservation, and compaction — from simplest to most powerful. Each involves a different trade-off between implementation complexity, information loss, and cost. Most production agents use a blend of all three, and the lecture walks through how OpenClaw, a widely-deployed open-source agent, implements compaction in production using "silent turns."

Read the full lecture narrative

Additional Resources

• Lecture slides
• OpenClaw Architecture, Explained — Full walkthrough of OpenClaw's runtime architecture, including its auto-compaction and memory systems.
• Deep Dive: How OpenClaw's Memory System Works — Technical deep dive into OpenClaw's silent turn, token thresholds, and MemoryIndexManager internals.
• Prompt Caching — Anthropic Docs — How to cache repeated prompt prefixes to reduce cost and latency — a complementary technique to compaction for stable context segments.
• Messages API Reference — Anthropic Docs — The canonical reference for the API calls used in the compaction demo.
• Compaction — Microsoft Agent Framework — Official Microsoft documentation covering sliding-window truncation, LLM summarization, and tool-result compaction strategies with Python and C# examples.
• Cutting Through the Noise: Smarter Context Management for LLM-Powered Agents — JetBrains Research empirical comparison of observation masking vs. LLM summarization on real agent tasks, finding masking was 52% cheaper and achieved better task solve rates.