Back to full roadmap
topiccore
Summary Buffer Memory
Periodically summarize old messages, keep the summary, drop the detail. Prevent context blow-up.
2 hours1 resources1 prereqs
LangChain's classic ConversationSummaryBufferMemory pattern:
- Last N messages kept raw
- Older than N compressed into one summary
- The summary evolves as the conversation grows
Trade-off: detail loss vs context savings. Practical: summarize every 10 turns in active chats, summary 200-500 tokens.
LangGraph approach: MessagesState + post-processing node auto-updates summary every N steps.