Memory-Augmented agents maintain continuity across conversations by persisting knowledge beyond a single session. At startup, the agent loads all memories from a JSON file and runs a relevance filter — using an LLM to select which memories are pertinent to the current question. This prevents context bloat from irrelevant history.
Three memory types reflect different kinds of knowledge: **episodic** memories (past interactions and their outcomes), **semantic** memories (facts learned about the domain), and **procedural** memories (learned strategies and preferences). Each type is tagged so the retrieval filter can reason about relevance by type.
After responding, the `consolidate` node extracts new learnings from the interaction and appends them to the persistent store with a timestamp. Over time, the agent becomes increasingly personalized and effective for its specific user and domain — building genuine long-term memory.