Files
Oleg Maslov 0c3e2ead3b
Some checks failed
Installer Smoke / installer-smoke (push) Has been cancelled
moving
2026-09-02 10:10:29 +02:00

2.9 KiB
Raw Permalink Blame History

Hive Mind — LoCoMo SOTA results index (single source of truth)

Tracked, in-repo, reproducible. Headline = 86.49% (7-lane W4, N=1540, same-judge vs Memori). Verify: node recount.mjs. Supersedes the non-reproducible 87.66% (see correction note below).

Headline: On LoCoMo under Memori's own protocol (gpt-4.1-mini answerer + judge, N=1540), Hive Mind = 86.49% overall, +4.54pp over Memori 81.95 (z=4.64, p<10⁻⁵), leading/tying every category. Two same-judge head-to-heads anchor it (Memori + Mem0). Fully local substrate.

Correction note (2026-07-01)

The prior headline 87.66% (2026-06-11 report) does not reproduce. Fresh 7-lane W4 + fresh judge = 85.19% on its own archived 2026-06-11 substrate, 86.49% on the current substrate. Cause: stale-verdict-replay bug (harness note 2026-06-15) inflated the original judge pass; that 1350-correct judgment set is lost. Adopted 86.49% as canonical (founder decision 2026-07-01).

The numbers (gpt-4.1-mini answerer+judge, N=1540)

7-lane W4 vs Memori — same-judge

Category Ours Memori (pub) Δ
single-hop 92.27 87.87 +4.40
multi-hop 80.50 72.70 +7.80
temporal 81.62 80.37 +1.25
open-ended 69.79 63.54 +6.25
overall 86.49 81.95 +4.54 (z=4.64, p<10⁻⁵)

Mem0 — same-judge, our ruler

Ours 86.49 vs Mem0 73.96 = +12.53pp overall; temporal 81.62 vs 50.78 = +30.84pp (write-time dating vs ingestion-time). Mem0 judgments 1139/1540 committed.

Token efficiency

Prior Config-D knee (28% tokens for ~1pp) was measured on an earlier answer set — re-measure on the current substrate before re-citing. Do not carry the old 87.66-anchored Pareto numbers.

Where everything lives

Committed + reproducible in THIS monorepo

  • benchmarks/results/locomo-sota-2026-06/this dir: report + INDEX + recount.mjs + data/{answers,judgments}/…N1540.jsonl (the pinned raw evidence, recount = 1332/1540).
  • Substrate code: packages/hive-mind-core/src/mind/{inprocess-reranker,search,resolve-relative-date,parse-date-window,raw-detail-lane,recall-context}.ts.

Reproduction harness (hive-mind-test/scripts/locomo)

40-cell-retrieval-gpt41mini.mjs (7-lane: PROFILES/DATEWIN/EPISODIC/RAWDETAIL), 41-judge-memori-gpt41mini.mjs, 42-report-memori.mjs. Always fresh OUT_TAG to avoid stale-verdict replay.

OSS public (marolinik/hive-mind)

Substrate code @ bc4eba1 (PR #14). Benchmark results being updated 73.1% → 86.49% (this arc).

Gotchas (cost real time)

  • Stale-verdict replay: 41-judge resumes by row count on the OUT_FILE → reusing a judgments file replays old verdicts. This inflated 87.66. Use a fresh OUT_TAG every run.
  • 42-report tokens/query line is a display bug; real tokens = context_tokens in answer rows.
  • Mem0 shares ~/.mem0/migrations_qdrant lock → run per-conv.
  • Zep + LangMem stay literature-sourced (founder 2026-06-16).