Evidence · measured, not claimed
Tested and measured.
CrossTalk was not built for any single benchmark. The isolation and recall numbers below come from open-source components and a harness you can re-run — locally or on your own cloud.
Isolation · cross-tenant recall rate (lower = better)
0.00%
CrossTalk · schema + RLS + keys
mem0-class filter100%
another filter100%
CrossTalk0%
Recall · LongMemEval (higher = better)
45.0%
Config A · raw-recall floor (n=100, published harness)
CONFIG B · IN PROGRESS
The write-path score (extraction, dedup, temporal supersession) lands with Config B. We publish the floor now rather than wait — and never a number we can't reproduce.
What these benchmarks test
- Cross-tenant leakage
- Whether one tenant's memory can surface in another's recall under entity-merge pressure.
- Long-conversation recall
- Whether facts scattered across a long, topically diverse history are retrieved on demand.
- Temporal supersession
- Whether a superseded fact is retired so recall returns current truth, not stale answers.
- Reproducibility
- Every result ships with an eval pack + code; run it locally or on your own cloud.