Rotate to Landscape

The benchmark registry is a dense evidence table and is locked to landscape view on mobile.

Benchmark Library
All Evidence runs · SHA256 verified artifacts · Verified Output recall

Completed 1 Million Turns in 541 hours


+ Software Comparison

1,015,114 verified turns · nuScenes AV · 1000-turn MemO(1) vs Mem0 · Dell OptiPlex 9020

2 panels

Final stopped run · Dell OptiPlex 9020 · RTX 3050 6GB · qwen2.5:7b

1,015,114 verified turns

541 hours 39 minutes on the nuScenes AV dataset with every turn SHA256 verified, zero failed turns, and a clean blind test result before retrieval.

Pass / fail

1,015,114 / 0

passes matched turns at stop

Recall latency

0.08 ms

MemO(1) memory path on the final snapshot

Model latency

768 ms

qwen2.5:7b response on the final snapshot

Active tokens

278

per verified turn

Open final dashboard Frozen status JSON YouTube benchmark video pending

MemO(1)™ vs Mem0 · server Qdrant · local Qwen 2.5 7B

1000-turn software benchmark, same machine, same queries

This is software protocol validation only. It does not claim FPGA timing. The run used 100 retrieval queries per history depth, 5 retrieval repeats, and 10 model-answer queries per depth.

Task MemO(1) retrieval Mem0 retrieval MemO(1) answer Mem0 answer Plain English
Exact recall 100% at 0.3287 ms p95 1% at 286.9869 ms p95 100% · 220.3 avg Qwen tokens 10% · 781.7 avg Qwen tokens MemO(1) found the exact record. Mem0 mostly missed and used 3.55x the prompt tokens.
Semantic recall 100% at 0.9329 ms p95 13% top-1, 25% top-5 at 351.96 ms p95 100% · 221.1 avg Qwen tokens 50% · 699.4 avg Qwen tokens Mem0 improved on fuzzy queries, but still used 3.16x the prompt tokens and answered correctly half the time.

Measured caveat: semantic MemO(1) uses a deterministic structured router before verified recall. Deterministic timing claims remain scoped to the FPGA lane only.