MemO(1) vs Mem0 · Public Evidence Report
Created UTC: 2026-08-20T19:23:08.223108+00:00

Scope:
Software validation only. FPGA and hardware timing were not touched in this batch.
Backend implementation details are intentionally withheld.

Run shape:
500 retrieval queries per condition.
5 retrieval repeats.
50 answer queries per condition.
5 answer repeats.
Local Qwen 2.5 7B answer model.
Mem0 used server Qdrant on the same workstation.

1000-turn exact recall:
MemO(1): 100% retrieval accuracy · 0.24 ms p95.
Mem0: 0.2% retrieval accuracy · 152.8548 ms p95.

1000-turn exact answer:
MemO(1): 98% answer accuracy · 217.06 average Qwen prompt tokens.
Mem0: 6% answer accuracy · 753.94 average Qwen prompt tokens.

1000-turn structured routing vs embedding retrieval:
MemO(1): 100% retrieval accuracy · 0.6548 ms p95.
Mem0: 10.2% top-1 retrieval accuracy · 28% top-5 · 189.241 ms p95.

1000-turn structured routing vs embedding answer:
MemO(1): 98% answer accuracy · 218.86 average Qwen prompt tokens.
Mem0: 30% answer accuracy · 696.4 average Qwen prompt tokens.

Plain English:
MemO(1) kept exact structured telemetry recall near-instant and bounded.
Mem0 improved on fuzzy queries, but it failed exact structured recall at 1000-turn depth in this workload.

Boundary:
This is not a semantic-search head-to-head claim. The fuzzy-query track compares structured routing against embedding search on telemetry questions where stable structure exists.
Deterministic timing claims remain FPGA-only.