Benchmarks

Five benchmarks, first on three

With one open model writing memory for every system, Reglos leads EverOS, MemOS and Mem0 on PersonaMem-v2, LongMemEval-S and LoCoMo.

Benchmarks

Results

Qwen3.8-27B answering

System	PersonaMem-v2	LongMemEval-S	LoCoMo	BEAM-100K	BEAM-500K
Reglos	61.3	84.0	87.2	67.9	65.6
EverOS	53.6	80.4	81.5	69.5	58.3
MemOS	53.8	73.0	75.7	67.0	62.9
Mem0	45.0	65.6	57.8	69.4	66.8

GPT-5.6 Luna answering

System	PersonaMem-v2	LongMemEval-S	LoCoMo	BEAM-100K	BEAM-500K
Reglos	57.5	87.2	89.5	65.9	65.0
EverOS	49.6	80.0	86.1	67.5	57.0
MemOS	50.5	74.0	78.8	64.4	61.3
Mem0	45.2	64.4	62.4	68.8	59.9

91.4%

Questions involving sensitive information

PersonaMem-v2 · other memory systems 42.9%–47.6%

63.1%

Things the user asked to forget

PersonaMem-v2 · EverOS 42.8%, Mem0 11.8%

98.2%

What the assistant itself said earlier

LongMemEval-S · EverOS 67.9%

+6.5

Gained by four self-improvement steps

54.2% → 60.7% · accepted on held-out users

Accuracy holds as more is written: writing only 25%, 50%, 75% or 100% of each history and asking the same questions, Reglos leads at every step.