Building a Sustainable Agent Memory Layer Achieving Recall 0.89 Using Elasticsearch

The context window of LLM agents only serves the role of short-term memory, and has limitations as a Long-term Memory system that persists even after sessions end. To solve this, a sustainable memory layer based on Elasticsearch was built.

This architecture is designed based on three categories from cognitive science, and includes the following core technologies:

• Hybrid Recall: Combines RRF (Reciprocal Rank Fusion) with a Cross-encoder reranker to improve search accuracy
• Supersession: A mechanism that updates to the latest information when contradictions occur between pieces of information
• DLS (Data Level Security): Maintains security through per-user memory isolation

In a QA-style evaluation consisting of 168 questions, an average Recall of 0.89 based on R@10 was recorded, demonstrating stable performance without data leakage in a multi-tenant environment.