Leaderboard & public results policy

July 15, 2026 · View on GitHub

FluctlightDB does not publish headline benchmark numbers to a third-party agent-memory leaderboard today. This page explains why, what we do instead, and what would change that.

Agent-memory “leaderboards” (LoCoMo, LongMemEval)

FactDetail
Canonical live leaderboard?No — LoCoMo and LongMemEval have official datasets/papers, but no neutral site where vendors submit scores and get ranked like BEIR or MTEB
What Mem0/Zep/etc. publishMostly self-reported blog posts, PDF tables, or GitHub README percentages
What we publishFrozen JSON in-repo + open reproduce scripts + honest REPRODUCIBILITY.md
Third-party submission statusNot submitted anywhere external for LoCoMo / LongMemEval

Our public numbers live here:

ArtifactPurpose
benchmarks/results/paper-2026-07-09.jsonPaper freeze bundle
benchmarks/results/locomo-invented-stack-engine-2026-07-13.json (MiniLM), locomo-mpnet-engine-2026-07-15.json (mpnet)LoCoMo cert — honest raw evidence recall, no expansion (96.8% @150 MiniLM / 97.0% mpnet, @5=72.6%/75.1% tight-k). The old 99.0% headline was ±3 neighbor-expansion inflation, deprecated.
make reproduce-locomoAnyone can verify LoCoMo cert locally
GitHub Releases + PyPIVersioned engine, not a ranking

What we are not doing yet (on purpose)

ActionStatusWhy
Mem0/Zep head-to-head blog postBlockedCredible only after first external LoCoMo reproduction (REPRODUCIBILITY.md)
Vendor roundup outreach (EverMind-style)Not startedWaiting on independent repro of headline numbers
Claiming “#1 on LoCoMo leaderboard”NeverThat leaderboard does not exist as a neutral registry

What we can submit (optional, different metrics)

VenueMetric typeFit
BEIR leaderboardIR nDCG@10SciFact 0.645 — standard IR, not agent E2E QA
arXiv / ZenodoPaper + frozen JSONAlready done (DOI on README)
Hugging Face dataset cardReproduce scriptsOptional future
Independent repro issue“Verified by X”Preferred — public credit in REPRODUCIBILITY.md

When we will pursue external visibility

  1. First external LoCoMo match → link from README + optional comparison post on identical protocol
  2. Independent LoCoMo/LongMemEval repro → add to REPRODUCIBILITY table
  3. Co-maintainer or named reviewer → stronger bus-factor story for adoption posts

Summary

We are not withholding from a leaderboard — there is no neutral agent-memory leaderboard to submit to.
We publish reproducible artifacts instead of registry entries. External credibility comes from independent reproduction, not from us self-listing on a vendor comparison site.

See also: MAINTAINER.md · BENCHMARKS.md