RippleMem: From Isolated Retrieval to Associative Recollection for Long-Term Agent Memory

TL;DR - RippleMem is a long-term agent memory system that retrieves initial “anchor” memories, then follows semantic and structural associations to reconstruct distributed evidence. It improves benchmark accuracy while cutting memory-graph construction cost by roughly 30×.

• Stores interactions as cue-rich episodic units in an event-centric memory graph.
• Uses hybrid cues for anchor retrieval, followed by adaptive associative expansion.
• Improves LLM-as-a-Judge accuracy by 3.95% on LoCoMo and up to 11.87% on LongMemEval-S.
• Addresses incomplete flat retrieval without relying on noisy full-context searches or costly graph construction.