Memory Summarization
You cannot keep 80 turns. A summary is the only way to stay coherent without the full log.
Meeting minutes — not a recording of every cough.
Visual Workflows
Start here — scroll inside each diagram frame to explore, then use + / − to zoom up to 200% if needed.
Overview
Scroll inside the frame to explore · use + / − to zoom up to 200%
Scroll inside the frame to explore · use + / − to zoom up to 200%
Rolling summary
Scroll inside the frame to explore · use + / − to zoom up to 200%
Old block → summary. Recent turns stay verbatim.
Key Takeaways
- 1.Summarization rewrites old context into a shorter running note. Update every N turns so the summary does not rot.
- 2.A bad summary is a false memory — the model will trust it. Keep a pointer to the raw log for audits.
- 3.Run a cheaper model to summarize the oldest block. Instruct it to keep ids, decisions, and open questions.
- 4.Store the summary as a system or developer turn. Keep raw logs off the hot path.
Learn elsewhere
- →Memory Compression
- →Conversation Memory
Real Example
Scenario
After 30 turns of a migration chat, the summary is: 'Moving billing to v2. Blocker: tax table. Owner: Priya. Do not drop EU VAT.'
What you would do
In Agent Memory, apply Memory Summarization to this scenario: After 30 turns of a migration chat, the summary is: 'Moving billing to v2. Identify the inputs, run the technique, validate the output, and note one thing you would monitor in production.
Commands
Commands to Remember
Summarize every N turnsPreserve ids and blockersKeep raw logs for auditA summary can lie
Cheat Sheet
Quick recap
quick ref- •Minutes, not a tape
- •Rolling updates
- •Ids must survive
- •False memory is a bug
Common Mistakes
- ✕Summarizing so vaguely that blockers vanish
- ✕Using the same huge model for cheap summaries
- ✕Throwing away the raw log