========================================================================
Potluck Memory Retriever — LongMemEval S Results
========================================================================
  Dataset:          longmemeval_s (500 questions, ~48 sessions/instance with filler noise)
  Instances run:    500
  Retrieval cutoff: k=10
  Embedder:         BAAI/bge-small-en-v1.5 (ok)
  Wall time:        369.2s

AGGREGATE
----------------------------------------
  recall_any@10:  0.9800
  recall_all@10:  0.9380
  nDCG@10:        0.8847
  p50 latency:    12.9ms
  p95 latency:    39.8ms
  errors:         0

BY QUESTION TYPE (recall_any@10)
----------------------------------------
  knowledge-update                    1.0000  (n=78)
  multi-session                       1.0000  (n=133)
  single-session-assistant            0.9643  (n=56)
  single-session-preference           0.9333  (n=30)
  single-session-user                 0.9857  (n=70)
  temporal-reasoning                  0.9624  (n=133)

NOTE: This recorded run measures retrieval only, not generated-answer accuracy.
It was run June 2, 2026 on a Mac M2 Pro. It has not been rerun for release 0.1.6.
Application source and the benchmark harness are not publicly released yet.