Agentic AI Notebook
Context Engineering
Phase 6Module 5 of 15

Context Compaction

Long sessions cannot stay as raw turns. Compaction is how agents survive hour-long jobs.

Meeting minutes with action items, not a transcript.

Visual Workflows

Start here — scroll inside each diagram frame to explore, then use + / to zoom up to 200% if needed.

Overview

100%
Loading diagram...

Scroll inside the frame to explore · use + / − to zoom up to 200%

100%
Loading diagram...

Scroll inside the frame to explore · use + / − to zoom up to 200%

This turn

100%
Loading diagram...

Scroll inside the frame to explore · use + / − to zoom up to 200%

Meeting minutes with action items, not a transcript.

Key Takeaways

  • 1.Compaction rewrites an old thread into a shorter running state.
  • 2.It is summarization plus structured fields — not a vague paragraph.
  • 3.A bad compaction is a false memory.
  • 4.Long sessions cannot stay as raw turns.
  • 5.Compaction is how agents survive hour-long jobs.

Real Example

Scenario

Compaction: pack only what this turn needs.

What you would do

In Context Engineering, apply Context Compaction to this scenario: Compaction: pack only what this turn needs. Identify the inputs, run the technique, validate the output, and note one thing you would monitor in production.

Commands

Commands to Remember

  • Compaction rewrites an old thread into a shorter running state
  • It is summarization plus structured fields — not a vague paragraph
  • A bad compaction is a false memory
  • Keep a pointer to the raw log

Cheat Sheet

Quick recap

quick ref
  • Compaction
  • Budget
  • Don't dump

Common Mistakes

  • Skipping evaluation for Context Compaction before production
  • No logging or tracing around context compaction steps
  • Ignoring cost and latency implications