Beyond RAG：EverMemModel Achieves SOTA by Ingesting Entire Databases at Once

The EverMemModel has achieved SOTA performance both on retrieval task and QA task.

EverMind researchers

Nov 18, 2025

About 1 minutes to read

SOTA

The EverMemModel achieves a technological breakthrough by allowing users to input the entire retrieval database along with their query into the model, which then rapidly returns reference document IDs and answers.

Retrieval Task: On NQ320K (full text), it achieves a Recall@1 of 75.5. For the unseen test set, the Recall@1 metric reaches 66.49, ultimately achieving SOTA on both NQ320K leaderboards.

QA Task: The DSA method performs QA directly on contexts up to 7.1M in length without relying on Embedding retrieval. When compared to the RAG method based on Qwen3-Embedding-4B + Qwen3-4B-Instruct and the Gemini 2.5 Flash method, it outperforms both (the metric in the table is the LLM Judgment score for Gemini 2.5).

Retrieval Task: On NQ320K (full text), it achieves a Recall@1 of 75.5. For the unseen test set, the Recall@1 metric reaches 66.49, ultimately achieving SOTA on both NQ320K leaderboards.

You may also like these

Sep 4, 2026

Multi-round retrieval: letting the model decide when to stop searching

multi-round retrieval,agentic retrieval,fixed top-k,injection budget,core selection,EverOS,agent memory,LoCoMo,LongMemEval,EverMemBench,SubtleMemory,multi-hop retrieval,retrieval accuracy,prompt token cost