Session memory compaction

This notebook focuses on manual, proactive session memory management for conversational applications: build a compact “session memory” in the background so you can compact instantly when context starts getting tight.

Upstream notebook: misc/session_memory_compaction.ipynb

When to use this

You run a long-lived chat experience (coding assistant, support agent, writing tool).
You want compaction to feel instant (no “please wait while we summarize” moment).
You need more control than “automatic compaction” (what gets preserved, what gets dropped, when it runs).

What to focus on

Session memory prompt design: preserve goals, constraints, decisions, and “known facts”
Background memory refresh (threading) so compaction is ready on demand
Prompt caching to reduce background update cost

Related notebooks

Automatic compaction in agentic workflows: tool_use/automatic-context-compaction.ipynb
Memory patterns and safety boundaries: tool_use/memory_cookbook.ipynb

Run locally

make test-notebooks NOTEBOOK=misc/session_memory_compaction.ipynb

Automatic context compaction Memory & context management