Skip to content

LongMemEval-S — Hybrid Retrieval

Headline result — like-for-like win over the reference

Hybrid retrieval (BM25 + all-MiniLM-L6-v2 vector fused via Reciprocal Rank Fusion, k=60) reaches Recall@1 0.892 / Recall@5 0.986 / Recall@10 0.992 / MRR 0.931, beating the agentmemory BM25+Vector reference (Recall@5 0.952 / Recall@10 0.986 / MRR 0.882). Both use the same retrieval-recall metric family on the same dataset, so this comparison is like-for-like and real.

Full 500 questions, re-confirmed 2026-08-07, not a sample

This page is the complete LongMemEval-S question set, no sampling. It replaces a run that predated the #457 weighted fusion change and whose only post-#457 evidence was a 100-question sample (#530). Recall@1 moved 0.894 → 0.892 (one question of 500), Recall@5 and Recall@10 are unchanged, MRR moved 0.9317 → 0.9307. LongMemEval-S was never affected by the #514 scoring correction: its ground truth is session IDs, which the old rule already emitted on both branches.

Run date: 2026-08-07
Retrieval mode: hybrid
Ranking unit: session (gold-blind, first occurrence wins)
Python: 3.12.13
Platform: macOS-26.5.2-arm64-arm-64bit
Sample mode: stride
Seed: 0
Limit: all

Summary

Metric Value
Questions evaluated 500 / 500
Errors 0
Skipped 0
Recall@1 0.8920
Recall@5 0.9860
Recall@10 0.9920
MRR 0.9307
Latency p50 (ms) 57.01
Latency p95 (ms) 71.91

By question_type

question_type n Recall@1 Recall@5 Recall@10 MRR
knowledge-update 78 0.9359 0.9872 1.0000 0.9634
multi-session 133 0.9023 0.9925 0.9925 0.9373
single-session-assistant 56 1.0000 1.0000 1.0000 1.0000
single-session-preference 30 0.7000 0.9667 0.9667 0.8111
single-session-user 70 0.9000 0.9857 1.0000 0.9326
temporal-reasoning 133 0.8496 0.9774 0.9850 0.9019

Notes

  • Retrieval mode: hybrid — ContextAssembler.assemble() is the primary path; effective mode resolves to 'hybrid'.
  • Hybrid fuses BM25 (lexical) + vector (all-MiniLM-L6-v2, 384-dim, in-process numpy cosine) via Reciprocal Rank Fusion (k=60).
  • Ranking unit: session — Every retrieved record is collapsed to its session ID before scoring — gold and non-gold alike. The unit is fixed by the dataset's ground-truth granularity and resolved before retrieval, so the answer key affects only the final metric (issue #514).
  • LoCoMo: image-only turns skipped (text-only evaluation).

Reference Numbers

agentmemory BM25+Vector (all-MiniLM-L6-v2) on LongMemEval-S: - Recall@5: 95.2%, Recall@10: 98.6%, MRR: 88.2%

Popoto BM25-only baseline on LongMemEval-S (any-hit, #438): - Recall@5: 95.2%, Recall@10: 97.8%

This run used hybrid retrieval (BM25 + all-MiniLM-L6-v2 vector fused via RRF, k=60). Compare Recall@5/Recall@10 above against the BM25-only baseline and the agentmemory hybrid reference.