Skip to content
MemJudge
LeaderboardCompareRunHow scoring worksGitHub

Corrections policy

Last updated 19 October 2026

MemJudge aims to be fair to every system it tests. If you think a result is wrong, unfair or should not be published, tell us.

How to report

Email admin@memjudge.org or open an issue at github.com/memjudge/memjudge. Say which result (its link), what is wrong and what you would like done. Anyone can report; vendors of a tested system get priority. We aim to reply within 3 business days.

What we can do

  • Correct, when the setup was wrong: for example a connector that did not follow the vendor's documented configuration. We fix it and re-run, or invite the vendor to.
  • Label, when a result stands but has a known limit: for example "self-hosted open-source version" or "the service cannot take conversation dates".
  • Remove, when publishing breaks a law, a court order or the terms that bind the person who ran it. While a report is decided we may hide the disputed result.

Right of reply and records

A vendor can ask for its response to be linked next to its results. Every correction or removal is recorded in the public corrections log with its date and reason. Results are signed, so a corrected result is published as a new, labelled result rather than by editing the original. A low score alone is not a reason to remove a result.

How we publish

We publish first and tell the vendor the same day, with links to the result and the connector code; we do not wait for permission. Every run of a connector stays listed on its run page, so a re-run after a fix never hides the earlier result. The fairness rules we follow are in publishing results.

MemJudge is a free, non-commercial, vendor-neutral benchmark for AI memory systems. The code is Apache-2.0 and scores are published under CC BY 4.0; dataset text in the results keeps its own licence (LoCoMo is CC BY-NC 4.0, non-commercial). It is built by Bit Robotics Ltd, which makes Recalld, so every run says whether its submitter is affiliated with the system it tests.

Every result here is signed. How signing works and how to check a result yourself.

Terms of Use · Privacy notice · Corrections policy · admin@memjudge.org

Live: site 4c60e4d · results ffc8d78