Summary Buffer Memory

A hybrid memory that keeps recent messages word-for-word while summarizing older history. Think about how you remember a long phone call with a friend. The last few sentences? You recall them almost word-for-word. That's your short-term memory. The earlier parts? You remember the gist, not the exact words. That's your long-term memory. Summary Buffer Memory works the same way. It keeps a buffer zone of the most recent messages in their original form. Everything older gets compressed into a running summary. The LLM receives both pieces on every call: the summary for historical context, and the raw recent messages for precision. Pure buffer memory (storing every message) gives exact recall but costs more tokens with each turn. Pure summary memory (compressing everything) saves tokens but loses recent detail. Summary Buffer Memory combines both strengths. It's especially valuable for customer support bots, coaching agents, and project management assistants. These agents need to remember the full arc of a conversation while responding precisely to the latest message. The core engineer…