New Memory Retrieval System Achieves #1 on LongMemEval Using Gemini Flash
Key point
A new memory retrieval system achieved SOTA on the LongMemEval benchmark using Gemini Flash.
Details
An experimental memory retrieval system took the #1 spot on the LongMemEval benchmark, scoring 96.4% (top-50) using Gemini 3 Flash. This result surpasses existing models using Gemini 3 Pro (Mem0 94.8%, Honcho 92.6%, HydraDB 90.79%, etc.).
The research intentionally used the smaller Flash model in order to independently evaluate Retrieval Quality rather than the capability of the model itself.
Key Design Features:
- Query decomposition: Performs parallel retrieval targeting different information needs in order to capture all information required from multi-session questions.
- Temporal salience scoring: Scores candidates by combining semantic similarity, lexical precision, and temporal salience.
- Coherence re-ranking: Re-ranks results by considering coherence and temporal flow between memories before passing them to the answering model.
Performance by Category (top-50):
- Single-session (User): 98.6%
- Assistant: 100%
- Preferences: 96.7%
- Knowledge Update: 97.4%
- Multi-session: 94.0%
- Temporal Reasoning: 95.5%
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.