To coach behavior over months, Buffy Agent needs memory that lasts longer than a single chat session. That's why the system is built with three distinct memory layers.

Short-term conversational memory

Short-term memory keeps the most recent dialogue and context in a fast store, typically Redis. It powers follow-up questions like "move that to tomorrow" without re-explaining which task you mean.

Episodic event history

Episodic memory logs concrete events: when habits were completed, when tasks were finished or skipped, when reminders fired and how you responded. This gives the agent a factual history to reason about, instead of guessing.

Semantic understanding of your patterns

Over time, the agent turns raw events into semantic memory stored in a vector database: patterns like preferring deep work in the morning or typically skipping evening workouts after late meetings. This semantic layer is what enables more personalized, generative suggestions.