01 · Memory May 2026
State-of-the-Art in agent memory
- Recall@k
- LongMemEval
- Performance
Barb achieves 95% Recall@15 with aggregation while adding only ~720 tokens — a 99.4% context reduction. High recall with minimal overhead across all six LongMemEval categories.
Read full paper 〉 LongMemEval_s · 500 questions · 6 categories · with aggregation
State-of-the-art
Recall@k
SS – Assistant
100%
SS – User
97%
Knowledge Update
99%
Multi-session
93%
Temporal Reasoning
91%
SS – Preference
90%
Overall 86% · ~190 mean tokens · 99.8% context reduction Overall 91% · ~450 mean tokens · 99.6% context reduction Overall 95% · ~720 mean tokens · 99.4% context reduction