07-context-compaction

07: Context Compaction

Key Insight: Summarize the old, keep the recent, never hit the limit.

What's New vs 06

• estimate_tokens() — Heuristic token counting (chars / 4)
• compact_session() — Summarizes old messages, keeps recent ones verbatim
• Automatic compaction triggered when token count exceeds threshold
• Session file gets rewritten with compacted history

What You'll Learn

• Why context windows have limits and what happens when you hit them
• The split-summarize-merge compaction strategy
• Token estimation heuristics
• What information survives compaction and what gets lost

Prerequisites

• Completed module 06
• Anthropic API key and Telegram bot token in .env

How to Run

make 07-context-compaction
# or
uv run python 07-context-compaction/bot.py

What to Observe

1. Long conversations work — The bot handles conversations that would exceed the context window
2. Key facts survive — Information from early in the conversation is preserved in summaries
3. Recent context is verbatim — The last N messages are kept exactly as-is

Files

File	Lines	Description
bot.py	~350 lines	Bot with automatic context compaction

Next Module

Ready for persistent memory? -> 08: Long-Term Memory