Building AI memory that holds up at 10 million tokens
Preliminary internal results published August 2, 2026. These are answer scores from a document-search evaluation, not a 99% enterprise-data retrieval measurement or an independently verified comparison.
YouWo memory benchmark results
- LoCoMo accuracy
- 94.6/100
- LongMemEval accuracy
- 97.2/100
- BEAM 1M average score
- 68.1/100
- BEAM 10M average score
- 52.1/100
Benchmark history
August 2, 2026: one founder-supplied preliminary snapshot. No later run or reproducible YouWo adapter, configuration, and per-question outputs have been published here yet.
Method
Conversation histories were represented as searchable YouWo documents or notes. The system retrieved evidence for held-out questions, generated answers, and judged those answers against benchmark ground truth.
Original research
LoCoMo · LongMemEval · BEAM · MemoryAgentBench for future learning evaluation