Fact check request for Mem0's regulated-memory evaluation results

[ebeirne]

We executed Mem0 OSS in our open regulated-memory evaluation and would like
the Mem0 team to review its column before we promote the results more widely.

The evaluation tests five properties: stale-revision suppression,
point-in-time recall, provable erasure, lookahead protection, and historical
audit-state reconstruction. Lians runs through the same harness.

Mem0's results, methodology, and per-cell evidence are public:

• https://github.com/Lians-ai/Lians/blob/master/docs/regulated-eval-results.md
• https://github.com/Lians-ai/Lians/blob/master/agentmem/benchmarks/adapters/mem0_adapter.py

We used Mem0 OSS 2.0.11 with the documented OpenAI configuration. We pinned
gpt-4o-mini after reporting a default-model compatibility issue in #6085,
so the evaluation would not score Mem0 on a failed setup.

If we missed an API or configuration that changes a cell, please share it.
We will rerun it, publish the result wherever it lands, and credit the
correction. Adapter pull requests are also welcome.

Please respond by July 31, 2026 if possible. Corrections remain welcome after
that date.

[ebeirne]

Update: the comparison page is now live with this discussion linked as Mem0's public right of reply: https://www.lians.ai/compare/mem0

If the adapter, default configuration, capability assessment, or scoring is wrong, we will rerun it and correct the page. The request remains a methodology check, not a product leaderboard.